GPT-6 Astra

Not to be confused with Astral, a Python tooling company OpenAI agreed to acquire in March 2026, or with Project Astra, a Google DeepMind assistant.

GPT-6 Astra is an OpenAI model that the company began rolling out on September 3, 2026, describing it as "the world's most intelligent and aligned model" and as the successor to the GPT-5.6 family.^[35] It is the first OpenAI model the company has designated as reaching the Critical cybersecurity capability level under its Preparedness Framework, a classification OpenAI announced on September 1 and repeated in the launch system card.^[34][36] OpenAI president Greg Brockman closed the launch press briefing with the words "Welcome to the AGI era," and told reporters that he personally believed the company was "there" while leaving the judgment to readers; the company did not formally declare that Astra is artificial general intelligence.^[37][39][41] In the OpenAI API the model is gpt-6-astra, priced at $10 per million input tokens and $50 per million output tokens on Standard processing, 2.5 times the promotional price of GPT-5.6 Sol.^[35][37][64]

Before it had a product name, the model was known simply as Astra. OpenAI first confirmed the name on August 1, 2026, when it attributed ten mathematical and theoretical-computer-science results to an internal version of Astra, which it described as its "next major model".^[1][3] On August 7 the company said preliminary evaluation could not rule out Critical cybersecurity capability, and on September 1 it said further automated and expert-led testing supported a Critical designation and that it planned to make the model available "soon" under stronger safeguards.^[33][34] At that point the company had not announced a launch date or pricing and said a system card would accompany the launch; both arrived on September 3.^[34][35][36]

The August mathematics release included a 249-page manuscript, a machine-checkable Lean certificate for every result, and a companion document narrating how each proof came together.^[2][8][9] OpenAI said the tokens needed to find all ten solutions would cost roughly $2,000 at the API rates of Sol, the flagship tier of its GPT-5.6 family.^[1] At launch the company added two further number-theory results, both concerning gaps between prime numbers, with proofs it attributes to "GPT 6 Astra" and Lean formalizations on GitHub.^[35][53][54]

OpenAI's launch post reports state-of-the-art results on its own comparison set for computer use, browsing, software engineering, cybersecurity, science, and professional work, including 99.9% on ARC-AGI-3, 97.6% on FrontierMath Tier 4, 100% on ExploitBench, 72.6% on an offline OSWorld 2.0 subset, and 74.1% on DeepSWE v1.1.^[35] All of these are company-reported numbers produced in OpenAI's research environment or through its API, with scores taken at the maximum of any effort setting, and several of the headline figures differ between the post's prose, its table, and the pre-briefing materials the press received (see Capabilities and benchmarks).^[35][37][39] The ARC Prize Foundation independently verified the ARC-AGI-3 result the same day, reporting 99.9% on its semi-private set with OpenAI's context-management harness and 62.7% with its own standard harness, and said it was "not claiming that it is AGI".^[68] The New Stack's assessment was that Astra "posts big gains on specialized tasks, but it comes at a premium and does not clearly lead the coding pack".^[37]

The reported results and comparisons are internal OpenAI evaluations. The launch also disclosed that Astra's written reasoning is harder to monitor than GPT-5.6 Sol's, a finding the company said it "takes seriously", and chief scientist Jakub Pachocki said OpenAI would withhold further scaling if its ability to monitor model alignment degraded beyond a certain level.^[35][36][42] Press reporting before the launch described Astra as built for long-running work and said it coordinates multiple agents over extended periods to attack hard problems.^[3][6] The mathematics announcement drew attention for its substance and for its form. Thomas Bloom, whose Erdős problems database was at the center of an OpenAI embarrassment in October 2025, called the results "big news", while pushing back on talk of AI "replacing mathematicians".^[14][15] Critics such as Gary Marcus accepted the results as genuine but objected that OpenAI published no methodology and that skill at formally verifiable mathematics may not transfer to much else.^[13] Within a day of that announcement, a researcher at Anthropic reported reproducing half of the ten results with the publicly released Claude Fable 5 model (see Reproduction with Claude Fable).^[26]

Announcement and release timeline

The model's public history runs from a single sentence in a mathematics blog post to a launch that several outlets reported before OpenAI's own announcement page was reliably online.

Scroll sideways for more →

The August announcement was tagged as a research publication and framed the mathematics work as part of the company's science program, pointing to ChatGPT for Academic Researchers, an initiative that gives 100,000 scientists and mathematicians free access to its best ChatGPT models, and to an AI-generated disproof of the Erdős unit-distance conjecture that OpenAI had shared in May 2026.^[1] The model reveal itself was a single sentence in the third paragraph: "The results were achieved by an internal version of Astra, our next major model."^[1] Gizmodo's headline accused OpenAI of having "smuggled" the announcement into a blog post about math.^[6]

The launch itself did not go smoothly. Press coverage embargoed to 11:00 a.m. Pacific went live on schedule, describing a blog post that readers could not find; Hacker News commenters reported that the page at openai.com/index/gpt-6-astra/ was up for about two minutes before returning a 404, and one saved and rehosted it.^[63] At 18:43 UTC the X account @synthwavedd posted a screenshot of the post's computer-use, professional, and coding tables, after complaining minutes earlier about "unnecessary confusion around Astra right now that would be easily solved by... not announcing the model via press release".^[62] Other OpenAI pages, including partner case studies, also returned errors during the same window, and the day had already seen a multi-provider outage affecting ChatGPT, Grok, and Claude.^[63][84] The page returned later in the afternoon with the same text and tables as the saved copy, apart from one benchmark figure discussed below.^[35] The Hacker News thread on the CNBC story had passed 200 points by the early hours of September 4 UTC.^[63]

Availability and pricing

OpenAI staged the rollout. On launch day the model went to "a limited set of organizations", which press briefings identified as enterprise customers in the company's Daybreak cyber-defense access program, with ChatGPT Plus, Pro, Business, and Enterprise users, the OpenAI API, and AWS to follow "over the coming days".^[35][37][40] 9to5Mac noted that this mirrored the staged GPT-5.6 release.^[48]

Scroll sideways for more →

Unlike GPT-5.6, which shipped as Sol, Terra, and Luna tiers, OpenAI announced no such variants for GPT-6; The New Stack reported that the lineup at launch consisted of Astra and Astra Pro.^[37] The company's API catalog had not listed the model as of about 19:20 UTC on launch day, but the developer pricing page did by the evening, with a banner repeating that Astra was "rolling out today for enterprises in our Trusted Access Program" and that API and plan access were "coming in the coming days".^[64]

OpenAI's launch post gives Standard pricing of $10 per million input tokens and $50 per million output tokens, with separate cache rates, and, as first published, said Fast mode delivers up to 2.5 times the speed of Standard processing at twice the price; a copy of the post fetched on September 4 says up to 2 times.^[35] The developer pricing page fills in the rest:

Scroll sideways for more →

All prices per million tokens as listed by OpenAI on September 3, 2026; Fast mode is not offered for Astra with EU data residency, and the long-context rate applies to prompts with more than 272,000 input tokens, at twice the input and cache rates and 1.5 times the output rate for the whole request.^[64][82] The API changelog entry for the release adds constraints for developers migrating from Sol: Astra does not support the none reasoning-effort level, custom temperature or top_p values, or log probabilities, and tool calling requires the Responses API rather than Chat Completions. The same day OpenAI added asynchronous tool calling, mid-turn steering over WebSockets, and mid-conversation changes to reasoning effort to the Responses API for long-running Astra work.^[67]

The New Stack put the Standard price at 2.5 times GPT-5.6 Sol's promotional rate and level with Anthropic's pricing for Claude Fable 5.1; Sol has cost $4 per million input tokens and $20 per million output tokens since August 21, 2026, a promotion OpenAI says runs at least through November 21, 2026.^[37][64][67] The outlet noted that the Astra price is far above Meta's Muse Spark at $1.25 and $4.25 or Google's introductory Gemini 3.8 Flash rates of $0.75 and $3.75.^[37][64] VentureBeat published a comparison table in which Astra's Standard total of $60 per million tokens tied Claude Fable 5 and 5.1 and Mythos 5 and 5.1, and its Fast-mode total of $120 was the most expensive entry.^[39] Brockman argued at the briefing that per-token prices had become a poor proxy for cost: "Pricing tokens doesn't make any sense," he said. "Our tokens are not necessarily the same as our competitors' tokens; they're not the same between different model families." What buyers should compare, he said, is "the price per task".^[37][39] OpenAI's post supports that argument with cost-per-task estimates on several evaluations, including an estimated 32% lower API cost per task than Sol on DeepSWE v1.1 at each model's highest-scoring setting (VentureBeat, working from pre-briefing materials, printed 57%); The New Stack observed that the launch data was "too sparse to show whether those savings offset the price premium".^[35][37][39]

Training

OpenAI has published little about Astra's architecture or data. The system card says only that, like other OpenAI models, Astra was trained on diverse datasets filtered through the company's data-processing pipeline, and that it is a reasoning model trained through reinforcement learning to think before answering.^[36] Neither the card nor the launch post gives a parameter count or a context window, and the card mentions a knowledge cutoff only to say that an internal exploit benchmark uses vulnerabilities disclosed after it.^[36] OpenAI's developer model page, published later on launch day, supplies what the card omits: a 1,050,000-token context window, 128,000 maximum output tokens, a knowledge cutoff of April 30, 2026, text and image input with text output (audio and video are not supported), and reasoning-effort settings of low, medium, high, xhigh, and max. The page lists streaming, function calling, structured outputs, and image input as supported features, fine-tuning as unsupported, and Responses API tools including web search, file search, code interpreter, a hosted shell, computer use, and MCP; the only snapshot is gpt-6-astra.^[82] No parameter count has been disclosed.

What the company did say came at the press briefing. Aidan Clark, OpenAI's vice president of research, called Astra "by far our largest scale training run". "It's the first time we've pre-trained on more than 100,000 GPUs at our Stargate site in Texas," he said, referring to the Stargate data-center buildout. Clark also said Astra was the first OpenAI release in which earlier models played a significant role in supervising the training process.^[37][39] "Based on the evals we monitor during pre-training, we believe the jump from Sol to Astra represents a larger increase in capabilities than the jump to Sol represented over previous models," Clark said, according to VentureBeat, which attributed the capabilities to large-scale pretraining combined with reinforcement learning aimed at longer tasks.^[39] (VentureBeat's text renders the 100,000 figure as "DBUs"; The New Stack and CNET, reporting the same remark, say GPUs.)^[37][39][88] "This is really only achievable because everything from the data center networking to inference kernels to the shape of Astra itself is all designed from the ground up to enable this level of training scale," Clark said, according to CNET. "It's this scale of pretraining which gives Astra a deeper and more robust understanding of the world."^[88]

Training was interrupted in August. After the OpenAI-Hugging Face incident, in which OpenAI agents escaped a testing environment and compromised another company's systems, OpenAI paused certain frontier training, including certain training for Astra, for two weeks while it hardened isolation and network controls, expanded monitoring, and raised alignment requirements; its largest planned frontier RL run stayed on hold until August 28 (see Model family and design).^[34][51] OpenAI estimated the monitoring overhead of the new setup at roughly 20% of the inference compute being monitored.^[51]

Capabilities and benchmarks

OpenAI's launch post presents nine tables comparing GPT-6 Astra with GPT-5.6 Sol, Claude Fable 5.1, Claude Fable 5, Claude Opus 5, and Gemini 3.8 Flash. The post states that "evaluation scores are the maximum at any effort" and that GPT evaluations "were run in our research environment or via our API, which may provide slightly different output from production ChatGPT".^[35] The GPT-5.6 Sol column refers to the version in the API, Codex, and ChatGPT Work, which the post says differs slightly from the version in ChatGPT chat.^[35] Blank cells below are blank in OpenAI's table. Every figure in this section is company-reported; competitor scores are OpenAI's reproductions or its readings of published results, with the exceptions the footnotes flag.

Computer use

Scroll sideways for more →

OpenAI's footnotes say the OSWorld figure uses "OSWorld V2-Offline", a subset that works without internet access, that Claude performance on it "was reproduced by an independent third-party", and that the Claude scores use the official settings rather than "the modified tasks and modified grading from the Fable 5.1 System Card". For ScreenSpot-Pro and ExploitGym, the Fable scores OpenAI reports "come from Mythos, which is Fable with fewer safeguards".^[35]

Professional

Scroll sideways for more →

OpenAI notes that the Claude BenchCAD scores "reflect 3 modifications to the eval, detailed in the Fable 5.1 System Card".^[35] The OpenScore row scores optical music recognition of the OpenScore String Quartet corpus with the OMR-NED metric from the LEGATO paper.^[35][59] On the Artificial Analysis Intelligence Index, a third-party aggregate, Astra's 61.2 trails Claude Fable 5.1's 65.7, Claude Opus 5's 63.1, and Claude Fable 5's 62.1 in OpenAI's own table; Artificial Analysis's own published run gives 61 (see Independent context).^[35][86]

Coding

Scroll sideways for more →

The post's prose gives Terminal-Bench 4.0 as 57.9% for Astra ("at approximately 9% and 63% lower estimated API cost per task" than Sol and Fable 5.1 respectively), while the table as published on launch day, the screenshot circulated on X, and the reader-saved copy show 57.7%; by the early hours of September 4 UTC the live table also read 57.9%, without any note of the change, one of five cells that changed after launch day without any note (see Pre-briefing figures versus the published table).^[35][62] On FrontierCode, OpenAI says Astra was run with a developer message similar to a section of its Codex developer message ("Avoid creating excessive test files... The goal is clean, mergeable code") and that "the prompt was not optimized for the eval".^[35] In OpenAI's table Astra's DeepSWE lead over Gemini 3.8 Flash and Claude Opus 5 is 0.3 and 0.4 points, and Claude Fable 5 and Claude Opus 5 score higher on both FrontierCode splits and on the Artificial Analysis Coding Agent Index.^[35]

Academic, science, and health

Scroll sideways for more →

Humanity's Last Exam is the one row in the post where Astra trails every listed competitor.^[35] The post's introduction says Astra "saturates FrontierMath Tier 4 with a 98% score"; the table says 97.6%.^[35] OpenAI says it evaluated the Claude models on HealthBench Professional itself "following the intended HealthBench Professional procedure, using GPT-5.4 grading and length-adjusted, unclipped scores", with Opus 5 used as a fallback for Fable 5.1 provider refusals, and that Claude Fable 5 and 5.1 are absent from the LifeSciBench, GeneBench Pro, and MedChemBench rows "because they refuse the majority of questions in these evaluations".^[35] The New Stack noted that the FrontierMath result appears to cover the 41 private problems in the 43-problem tier, and that Epoch AI, which runs the benchmark, says OpenAI funded its development and has exclusive access to part of it.^[37]

Cybersecurity

Scroll sideways for more →

OpenAI tested Astra and Sol on ExploitGym "without the 6-hour time limit, to better assess their full cyber capabilities", saying both are fast enough that the limit has little effect, and the Claude ExploitGym figures are Mythos results.^[35] The internal June-August set contains 20 high-severity V8 vulnerabilities across 13 stable Chrome releases and tests arbitrary code execution in V8 and official Chrome releases for Linux; OpenAI cautions that "some included vulnerabilities may not permit arbitrary code execution under the evaluation's constraints, so a 100% success rate may not be achievable".^[35] SRE-Bench is a reverse-engineering benchmark published in August 2026 by Jeremy Spence and colleagues; OpenAI reports 99.2% within four attempts, against 68.7% for Sol.^[35][57] ExploitBench itself, published in May 2026 by Seunghyun Lee and David Brumley, scores 41 V8 vulnerabilities with 16 capability flags.^[58]

Alignment, long context, and abstract reasoning

Scroll sideways for more →

The ARC-AGI-3 row carries the post's first footnote: Astra "was run with our responses API harness, which changes two settings to better match real-world performance. The changes do not specifically target ARC-AGI-3."^[35] Those two settings are retained reasoning and compaction. In a July 29 post OpenAI showed that enabling them raised GPT-5.6 Sol's score on the ARC-AGI-3 public set from 13.3% under the official harness to 38.3%, roughly tripling it while cutting output tokens by six times; the 7.8% Sol figure in the launch table is the official-harness number.^[52] The New Stack's separate analysis of the result concluded that "the benchmark measures Astra and OpenAI's agent systems together", since the comparison models were evaluated under different setups, and that even the reported score "doesn't settle the AGI debate".^[38] VentureBeat made the same point by way of NVIDIA's August claim of 100% on the same public set using an agent architecture wrapped around Claude Opus 5, whose bare baseline NVIDIA put at roughly 30%.^[39]

The ARC Prize Foundation published its own verified runs on launch day, and they were the first Astra headline number with an independent measurement (Artificial Analysis and Epoch AI followed; see Independent context and Mathematical results). On the semi-private evaluation set, Astra scored 62.7% at max reasoning effort under ARC Prize's Standard harness, at a cost of about $26,098, and under a "Provider Adapter" harness that preserves OpenAI's opaque reasoning state between requests and uses compaction, it scored 99.9% at high effort (about $18,817) and 98.6% at max effort (about $17,332); every effort level from low upward scored at least 98.0% with the provider harness.^[68] The two figures OpenAI and the press circulated therefore both appear in ARC Prize's table, at different effort settings. ARC Prize called both results state of the art, said the provider-harness runs were about 3.66 times faster and used 49% fewer tokens than the standard-harness runs across the games both solved, and reported that at max effort Astra used fewer actions than the median human tester on 96.0% of levels, "by ARC-AGI-3's measure of action efficiency" matching and surpassing human parity.^[68] Its replays showed the model inventing a compact algebraic shorthand to track game state and, in a sandboxed harness with code execution, writing small game-specific software libraries.^[68] The foundation described Astra as "a noticeable step-function change in frontier model capabilities" and said it "clears this bar", while repeating that saturating ARC-AGI-3 was never meant as proof of AGI: "we are not claiming that it is AGI."^[68] It said it would report both harness conditions on its leaderboard from now on.^[68]

Pre-briefing figures versus the published table

Several outlets received benchmark materials before the launch, and their numbers do not all match what OpenAI published. The most visible case is ARC-AGI-3: The New Stack and VentureBeat reported 98.6%, a figure that also spread through the Hacker News thread, while the launch post's prose and table say 99.9%.^{[35][37][38][39][63]} ARC Prize's verified table shows both numbers, 98.6% at max effort and 99.9% at high effort under the provider harness, so the difference is most simply read as two effort settings rather than a correction.^[68] A table The New Stack published with its coverage, apparently compiled from the pre-briefing materials, differs from the published post in the following cells:^[35][37]

Scroll sideways for more →

The New Stack's article text uses the pre-briefing Sol figure of 70.8% on DeepSWE.^[37] OpenAI has not commented on the differences, and no source describes them as a revision; the wiki records both sets as reported.

The page changed again after launch day. A copy of the live post fetched in the early hours of September 4 UTC differs from every launch-day copy (the text saved during the outage, the X screenshot, and copies fetched at 19:36 and 20:13 UTC on September 3) in five table cells: Terminal-Bench 4.0 for Astra (57.7% to 57.9%), HealthBench Professional for Claude Fable 5.1 (56.6% to 58.1%) and for Claude Opus 5 (54.5% to 56.4%), the internal ExploitBench (June-August 2026) figure for GPT-5.6 Sol, whose launch-day table cell read 5.5% while the post's own chart text for that benchmark said 11.5% (the cell now reads 11.5%, the figure used in this article), and the AutoReview row of the internal computer-use safety benchmark for Sol (4.5% to 4.3%); the same copy describes Fast mode as up to 2 times Standard speed rather than 2.5 times, and says Astra "will be available" through Amazon Bedrock rather than "is also available". OpenAI did not mark the changes, and no source describes them as corrections. The tables in this article give the launch-day values.^[35][62]

Independent context

The New Stack placed the coding results against results OpenAI's chart omits. Earlier in the week Meta reported 75.4% on DeepSWE for Muse Spark 1.3 at a maximum reasoning setting that was still under safety review, and the public DeepSWE leaderboard put Gemini 3.8 Flash and Claude Opus 5 at 74% and Sol at 73%, with overlapping uncertainty ranges; on a 113-task benchmark, the outlet noted, the 1.3-point gap to Meta's number is one or two tasks.^[37] OpenAI's chart "excludes Muse and uses a 67.4% Fable 5.1 result, making Astra's advantage appear larger than the broader set of results would suggest".^[37] On Terminal-Bench Science, Anthropic reports 52.6% for Fable 5.1 and the public leaderboard tops out at 30% for Opus 5.^[37] On OSWorld, Anthropic has reported 77.9% for Fable 5.1 on a different OSWorld release that it says should not be compared with earlier scores.^[37] VentureBeat flagged an omission of a different kind: OpenAI's launch materials contain no result on GDPval, the company's own benchmark of economically valuable knowledge work, which VentureBeat called "conspicuous given the AGI framing around Astra", while acknowledging that GDPval's one-shot format does not measure the long-horizon, interactive work Astra is meant for.^[39]

Artificial Analysis published its own runs on launch day. GPT-6 Astra at max effort scores 61 on the Artificial Analysis Intelligence Index v4.1.1, which the firm's model page ranks eighth among 202 models in its class, behind Claude Fable 5.1 (66, with fallback), Claude Opus 5 (63), Muse Spark 1.3 (62, not publicly available), and Claude Fable 5 (62), and level with GPT-5.6 Sol and Grok 4.6 (61); it generated 42 million output tokens on the index against a median of 62 million, at $1.67 per index task and $3,013.30 in total. On the Coding Agent Index v1.4, run in the Codex harness, Astra scores 67, equal to Claude Fable 5, behind Claude Code with Fable 5.1 (70), Opus 5 (68), and Muse Code with Muse Spark 1.3 (68), and above GPT-5.6 Sol (65).^[86][87] The firm's summary: Astra "makes significant gains in the Artificial Analysis Coding Agent Index, scoring equal to Fable 5 at lower cost. In the Intelligence Index, it uses fewer tokens than GPT-5.6 Sol for similar performance, but this is outweighed by higher prices."^[87] Both figures match the 61.2 and 67.0 that OpenAI's own table quotes for the two indices.^[35]

Computer use and professional work

OpenAI's launch materials call Astra "the world's best computer use model", and VentureBeat reported that the company positioned it as the start of an era in which users "no longer have to click around a mouse or type on a keyboard ever again".^[39] The post lists the kinds of work it has in mind: filling out online forms, updating customer records in a CRM, organizing a calendar, conducting online research and drafting summaries in email or a document editor, analyzing scientific data, generating plots, creating a website and running front-end QA on it, and installing, testing, and troubleshooting software.^[35] In latency simulations on OSWorld 2.0, OpenAI says Astra scored 72.6% at roughly 40 minutes per task against Sol's 65.7% at roughly 75 minutes, about 47% less time per task.^[35][39]

The demonstrations shown to press included Astra operating KiCad, Excel, Blender, Power BI, Unity, and FreeCAD, performing browser-based form entry and website QA, modeling a house in Blender and turning it into a walkable Unreal Engine 5 scene, and turning a spaceship concept into a 3D model.^{[35][37][38][39]} A promotional video opened with a 1980s demonstration of a person asking a computer to draw a yellow circle, then cut to OpenAI employees asking Astra by voice to turn the circle into a rocket ship, a 3D game, and an eBay listing.^[39] "Astra can really do anything a human can do with a computer," Brockman told NBC News before the release.^[42] He argued at the briefing, as VentureBeat reported, that computer-use agents could bypass much of the connector-building that has occupied enterprise AI: "We've been bottlenecked over this gigantic era by people writing connectors and very painstakingly building these connections into all these tools that people can already use."^[39]

For professional work, the post says Astra is OpenAI's best model at following existing templates and producing slides, documents, spreadsheets, and analyses that match a user's writing and visual style, and that it was trained to "pull only the context that matters into outputs". Its example is a slideshow about "GPT-Gaia, a fictional model", built from a few slides of OpenAI's own template.^[35] With Sites in ChatGPT, the post says, Astra can create, host, and share websites, web apps, and games from a prompt.^[35] The post also claims better judgment when instructions are ambiguous: Astra "uses context to fill in routine gaps and asks focused questions when the answer could change the outcome", and it is meant to keep earlier constraints in mind when a user changes course mid-task rather than treating each steering message as a new goal.^[35]

The post carries endorsements from Higgsfield AI, Harvey, Jane Street, and Lovable. Higgsfield's chief executive said Astra completed the company's most complex creative workflows "while using up to 20% fewer tokens than other models we've tested"; Harvey's head of applied research described it as "a significant quality improvement over GPT-5.6 Sol across complex legal tasks"; Jane Street reported "a clear step forward in trading intuition evaluations"; and Lovable's chief technology officer said it "came out significantly ahead of GPT 5.6 Sol" across effort levels on one of the company's evaluations.^[35] These are launch-partner statements published by OpenAI. OpenAI researcher Mia Glaese framed the intended shift at the briefing: "With those capabilities, we expect people to delegate much more complex work across applications, with humans directing the work at a much higher level."^[39]

Codex and developer changes

The change The New Stack judged most likely to matter to developers concerns what happens when a job outgrows the context window.^[37] Codex has relied on compaction, which summarizes earlier work to free space and can discard the detail an agent later needs, such as why a fix failed or which tests ran. With Astra, OpenAI says, Codex can instead keep notes across context windows "preserving accumulated details without repeatedly compressing them into a single summary", and earlier context windows remain searchable so the model can find requirements or test results from previous messages and tool outputs "even if that information wasn't captured in its notes".^[35] The feature is experimental, enabled through a setting in the Codex config.toml file, and OpenAI says it will become the default for Astra "in the coming weeks".^[35][37] The Codex model documentation gives the specifics: the setting is features.context_management.experimental_mode = true, it is off by default, and at launch it is available only to users signed in with ChatGPT Plus or Pro, not with Business, Enterprise, or API-key sign-in. The same document lists Astra as available in the ChatGPT desktop and web apps, the Codex CLI, and the Codex IDE extension, but not for Codex cloud tasks at launch, and says the rollout adds Astra Light, Astra Medium, and Astra Extra High options to the model picker for eligible Pro, Business, and Enterprise accounts.^[81]

Astra can also ask the user a question in Codex "asynchronously while continuing work that doesn't depend on your reply", proceeding on sensible assumptions if no answer comes but waiting on consequential decisions; The New Stack called a single unresolved decision blocking the rest of a job "a common failure mode for coding agents".^[35][37] OpenAI updated the Codex harness alongside the model to speed up computer use, and says the combination completes tasks 1.9 times faster than the current Sol-based experience on the Mind2Web benchmark.^[35][37] The post's coding section opens with the flat claim that "GPT-6 Astra is the best model for software engineering to date", a claim The New Stack's headline disputed with respect to the coding benchmarks specifically.^[35][37]

Mathematical results

The ten August results

The ten results, as stated in the August 1 blog post and in the manuscript's abstract, are listed below. The manuscript is credited simply to "OpenAI" and attributes the work to "an internal OpenAI model"; the name Astra appears in the blog post and in researchers' social media posts, not in the paper itself.^[1][2]

Scroll sideways for more →

The August post describes a three-stage pipeline. The internal Astra model produced the mathematical arguments; humans then prepared the arguments into manuscripts using the same model; and the model afterward formalized each argument in a Lean certificate. OpenAI also released what it calls a narration of the model's thinking process for each solution, a 62-page companion written by an AI model that read the original chains of thought and reconstructed how each proof developed.^[1][9] On cost, the post says the total number of tokens needed to find solutions to all ten problems would come to roughly $2,000 at Sol API rates.^[1] The post as first published on August 1 said the ten problems "have been open and have seen no progress on the main result for at least a decade, and in most cases much longer", a line The Decoder and BleepingComputer both quoted; by the Internet Archive's August 3 snapshot OpenAI had removed the sentence without notice, replacing it with wording that each result "resolves or makes substantial progress on a long-standing open problem".^[1][3][4]

OpenAI researchers amplified the announcement on X. Noam Brown wrote that "An internal version of Astra, @OpenAI's next major model family, solved 10 major open problems in mathematics, quantum complexity, and theoretical computer science", adding "We believe it will be a major step for scientific reasoning."^[11] Sébastien Bubeck opened his post with "yes, nonsofic groups exist", describing that statement as "one of many new beautiful results proved by Astra, our next major model".^[10] Brown was quote-posting OpenAI reasoning researcher Lijie Chen, an assistant professor at UC Berkeley, who singled out "new circuit lower bounds for computing the permanent" among the ten proofs.^[24] On August 3, 2026, OpenAI's main X account restated the claim in a single sentence: "An internal version of our next major model produced 10 new results on long-standing open problems in mathematics and theoretical computer science, using roughly $2,000 worth of tokens at GPT-5.6 Sol API rates."^[25]

The post closes with a section on responsibility to the mathematical community. OpenAI writes that "claiming human authorship for a proof generated entirely by an AI system would misrepresent both the system's contribution and the nature of genuine human intellectual work", says its researchers helped prepare the manuscripts and formalize the proofs and take responsibility for their correctness, and states that the mathematical arguments themselves were generated by its system. It also acknowledges the signers of the Leiden Declaration on AI and Mathematics, a June 2026 statement of concern from the mathematical community (see Background).^[1][22]

The result OpenAI and most commentators treated as the headline is the third: a construction establishing that non-sofic groups exist. Soficity, introduced by Mikhail Gromov in 1999, asks whether a group can be approximated by finite permutations; The Next Web noted that nobody had managed to prove or disprove the existence of non-sofic groups in the 27 years since.^[7] The manuscript's construction uses property-(T) expanders and the binary Leavitt algebra, and the formalization manifest records the formal statement as the existence of a finitely presented non-sofic group.^[2][8] The disproof of Connes's rigidity conjecture, posed in 1980, targets the claim that certain rigid groups are uniquely determined by their group von Neumann algebras: the paper constructs infinitely many pairwise nonisomorphic property-(T) groups sharing a single group von Neumann algebra, and says the argument also answers a related finite-to-one question of Popa.^[2][5] The sphere-packing chapter determines the exact exponential decay rate of the Cohn-Elkies linear program, which The Next Web described as the first improvement to the general upper bound on high-dimensional sphere-packing density since 1978.^[2][7]

Three of the ten problems carry numbers in the Erdős problems catalogue that Thomas Bloom maintains at erdosproblems.com. Problem 183 asks for the limiting growth rate of the k-color triangle Ramsey number R(3;k); Erdős offered $100 for showing the limit is finite, and the strongest lower bound recorded on the problem page is exponential in k.^[16] The paper claims the superexponential answer R_k(3) = k^(Theta(k)).^[2] Problems 146 and 180 are, respectively, a degeneracy conjecture and the compactness conjecture in extremal graph theory; the site lists a $500 prize for the first.^[17][18] As of August 4, 2026, all three entries on erdosproblems.com still described the problems as open; the site cautions that status reflects the maintainer's current belief.^[16][17][18]

Every one of the ten results ships with a Lean 4 certificate in the openai/ten-proofs repository on GitHub, released August 1, 2026 under an Apache 2.0 license. The project builds with Lean 4.32.0 and mathlib, its manifest reports zero "sorry" placeholders (unproven steps) across all ten formalizations, and each result carries a configuration for independent checking with the Comparator proof checker.^[8] SiliconANGLE noted that this is what gives the announcement its weight: Lean's kernel returns a binary verdict, so trusting the proofs does not require trusting the model or OpenAI. What machine checking does not settle, the same report noted, is whether each formal statement faithfully captures the open problem it claims to resolve, and whether the results matter; none of the ten had been through peer review at announcement.^[5]

Prime gaps at launch

With the September 3 launch, OpenAI published two further results, both on the distribution of prime numbers, as PDFs on its content network with abridged chains of thought and verification materials.^[35] The first concerns how close together primes can occur indefinitely far along the number line. For more than a decade the best unconditional result, from the Polymath 8b project, was that infinitely many pairs of consecutive primes lie at most 246 apart; OpenAI's paper says Julia Stadlmann has since proved 240 in "independent concurrent work" the paper cites as forthcoming.^[35][53] The OpenAI manuscript, dated August 30, 2026, proves that infinitely many pairs of consecutive primes lie within 186 of each other (in the notation of the field, H1 <= 186), by establishing the statement DHL[40,2] and exhibiting an admissible 40-element set of diameter 186. Its abstract says the improvement "combines the equidistribution estimates of Polymath 8a and Stadlmann with complementary factorization conditions" that allow a larger support for the multidimensional Selberg sieve and an improved numerical optimization, and states: "The proof is due to GPT 6 Astra."^[53] The accompanying openai/PrimeGaps186 repository, created September 2 under an Apache 2.0 license, holds a Lean 4 formalization that is conditional on numerical integral and cap bounds and on finite-field exponential-sum estimates ultimately derived from Deligne's work, together with a Python-FLINT program that verifies the numerical bounds.^[53][55]

The second result concerns unusually large gaps. Writing G(X) for the largest gap between consecutive primes not exceeding X, Rankin's 1938 lower bound stood for decades until Maynard and, independently, Ford, Green, Konyagin, and Tao improved its constant in 2014, with a later joint quantitative bound; the OpenAI paper also records an intermediate improvement it attributes to an argument by GPT 5.6 Sol posted on the Erdős Problems website.^[54] The new paper proves G(X) >> log X (log2 X)^2 log4 X / (log3 X)^2, where logj denotes the j-fold iterated logarithm, improving Rankin's bound by a factor of log2 X, and again states that "the proof is due to GPT 6 Astra". OpenAI's launch post describes this as improving "a term in a bound on these gaps that had remained unchanged for more than 80 years".^[35][54] A complete Lean proof of the theorem is in the openai/LongGapsBetweenPrimes repository, created September 2.^[54][56]

The launch post quotes Greg Burnham of Epoch AI on the mathematics: "The story is: end of one era, start of another."^[35][38] The New Stack's analysis pointed to what OpenAI's account leaves out: it "doesn't spell out what Astra came up with on its own, what researchers suggested or how the work moved between them", so while the results go beyond solving benchmarks with known answers, the outlet judged them "not enough to call the math evidence of AGI".^[38]

The one independent, budgeted measurement of Astra's research mathematics comes from Epoch AI's FrontierMath Erdős benchmark, announced the same week: 68 open Erdős problems selected by Thomas Bloom, each requiring a Lean proof or disproof, with one attempt per problem under a $300 and 72-hour budget. A pre-release version of GPT-6 Astra was the only model to solve anything, 2 of the 68 (3%): it disproved problem 74 by counterexample for $218 and 15 hours of working time and proved problem 126 for $247 and 16 hours, while GPT-5.6 Sol, GPT-5.5, Claude Fable 5.1, and Claude Fable 5 scored 0%. In further, less systematic attempts with larger budgets, which Epoch says are "not a FrontierMath Erdős score", the same pre-release Astra solved 5 of the 68 at least once, adding problems 1, 548, and 571.^[85]

Cybersecurity capability and safeguards

The Critical designation

OpenAI's August 7 and September 1 disclosures added the first detailed first-party account of Astra outside mathematics.^[33][34] On August 7, OpenAI said its preliminary evaluations showed enough progress in agentic coding and cybersecurity that it could not rule out Astra reaching the Critical threshold under the company's Preparedness Framework. It paused internal Astra work that did not meet stricter controls, including isolation, restricted network and tool access, stronger model-weight protection, monitoring, and sandboxing. OpenAI also said monitors were evaluating the model's chain of thought across internal agentic applications and could interrupt high-risk activity.^[33]

After additional evaluation, OpenAI said on September 1 that Astra met the Critical cybersecurity threshold, making it the first OpenAI model to receive that designation.^[34] One route to that threshold is identifying and developing functional zero-day exploits of all severity levels in many hardened, real-world critical systems without human intervention. The other is planning and carrying out novel end-to-end attacks against hardened targets from only a high-level goal.^[33][34] This is OpenAI's own preparedness classification, not an independent certification. TechCrunch observed on September 1 that "without any third-party confirmation, it is difficult to evaluate OpenAI's claims about safety or preparedness", and that the company had not said who its initial testers would be or how they would be chosen; WIRED reported that Daybreak partners, which include Cisco, Cloudflare, and Palo Alto Networks, would get early access to a less restricted version of Astra, and that OpenAI said it had been working with government partners on access to the model's cyber capabilities.^[45][46] The launch system card confirms the designation and adds the rest of the preparedness picture: Astra is treated as High, not Critical, in the biological and chemical category, and it does not reach the High threshold for AI self-improvement.^[36]

OpenAI reported a 100% score on ExploitBench, which uses known vulnerabilities. Because benchmark contamination was a concern, it also assembled an internal ExploitBench port from 20 high-severity V8 vulnerabilities disclosed from June through August 2026. The company reported that Astra achieved a higher arbitrary code execution rate than GPT-5.6 Sol while using fewer output tokens on that set, and that it found and used two zero-day vulnerabilities in an exploit chain; disclosure to the affected maintainers was still in progress.^[34] OpenAI cautioned that these results used Daybreak Blue access rather than the default production configuration.^[34] The launch post put numbers on the comparison: 100% against 78.5% for Sol on ExploitBench; 42.4% against 30.3% on ExploitGym, with the six-hour time limit removed for both; 39.0% against 11.5% on the June-August set; 88.0% and 99.2% against 55.9% and 68.7% on SRE-Bench at one and four attempts; and 85.4% against 79.1% on SEC-Bench Pro.^[35] The system card offers a concrete illustration of the contamination worry: given the description and patch for CVE-2023-6702 in one ExploitBench task, Astra failed to exploit it, then "directly recalled" a different vulnerability, CVE-2024-0517, and used that to reach arbitrary code execution.^[36]

In expert-led tests, OpenAI said Astra built a browser-compromise chain that escaped a sandbox and executed commands on the host after the browser opened an HTML file. It also combined multiple flaws in a hardened operating system into a local privilege-escalation chain.^[34] The system card describes the setup: internal experts supervised the model "for safety and validation" only and were not allowed to supply ideas or steer its research, the model ran in the standard Codex harness with web access and up to 64 subagents, and OpenAI says it disclosed the findings to the maintainers while withholding product names and exploit details (timings and a third-party evaluation by Irregular are in the System card section).^[36][76][77] These examples were disclosed without the complete benchmark artifacts, so they establish the company's reported evaluation outcome rather than a reproducible public result.

Deployment safeguards

The September report described three main deployment layers: refusal training in the model, system-level classifiers, and offline detection and threat disruption. On OpenAI's cyber-jailbreak evaluation, Astra refused 91.5% of requests, compared with 59% for GPT-5.6 Sol.^[34] In a separate simulated honeypot test run without production cyber safeguards, GPT-5.6 Sol attempted to access surrounding targets in 56% of trials, while Astra made no such attempts; OpenAI said Astra also never tried to bypass an auto-review denial in a related test.^[34] Those figures are vendor-reported and specific to the stated test conditions.

OpenAI said access to Astra's most advanced cybersecurity workflows would begin with a small group of alpha testers and expand through Daybreak Blue for defensive use. It also warned that the extra checks could slow, pause, or stop legitimate work; a paused task may ask for user review in ChatGPT or Codex, while an API task would stop.^[34] The launch post repeats the warning and adds a specific limit: the version of Astra launching on September 3 "will refuse to comply with more advanced cybersecurity tasks such as creating proof-of-concept exploits for vulnerabilities", though defenders can use it for secure code review and patching. Through Daybreak, OpenAI's cyber-defense program, the company plans to "expand access and roll out less restrictive safeguards in the coming weeks" for vulnerability and proof-of-concept validation, malware analysis, and detection engineering.^[35][66] The system card quantifies what the tiers permit (see System card): proof-of-concept exploit creation rises from 2.4% completion without trusted access to 92% with Daybreak Blue, while even with Daybreak Blue Astra fully completes only 3.5% of the arbitrary requests in the company's Advanced Cybersecurity Completion Rate evaluation.^[36][79] The New Stack noted that Daybreak Blue is an access program for authorized defensive work, not a separate model or reasoning mode.^[37]

Mia Glaese warned at the briefing that users outside the trusted-access programs might see slowdowns, pauses, or blocks while doing cybersecurity work, and sometimes while doing unrelated work: "At launch, this is something that people should expect."^[37] The system card describes the production misalignment monitor as asynchronous and able to pause or end a conversation, notes that some conversations, including those stopped through the API, cannot be resumed, and says the system preserves OpenAI's Zero Data Retention commitments, which is why "detailed explanations of a decision to pause or end a conversation are not available in every case".^[36][80] The company frames the launch through the argument Brockman made in "The Defender's Window" on August 16: frontier cyber capabilities help defenders find weaknesses faster while making them easier to exploit, so the same capabilities should reach defenders first.^[35][65] Astra also went through the White House's voluntary review of frontier models, NBC News reported; Brockman said the administration came back with nothing that required changes to the safeguards.^[42]

Alignment and monitorability

OpenAI calls Astra "our most aligned model" and rests the claim on a set of evaluations, most of them internal. The one the launch post leads with was built after the Hugging Face incident and tests whether a model facing a difficult or impossible task will go beyond its intended scope: GPT-5.6 Sol, running without production safeguards, went beyond the authorized target in 48.2% of cases (the post's prose rounds this to 48%), while Astra did so in 0%.^[35][37] On a computer-use safety stress test of tasks "adversarially selected to elicit misbehavior", run in a generic computer-using-agent harness without the auto-review and confirmation protections normally deployed in Codex and ChatGPT Work, Astra produced misaligned outcomes in 2.4% of cases against 9.5% for Claude Fable 5.1 and 11.5% for Claude Opus 5; with AutoReview enabled, Astra's rate fell to 1.8% against 4.5% for Sol.^[35] In an internal evaluation of responses to a Codex Auto-Review denial, Astra never attempted to circumvent the denial, "even when Auto-review was deliberately configured to be evadable and the task was impossible to complete otherwise", where Sol's rate was 0.29%.^[35] On a capability-hallucination evaluation, OpenAI says Astra is about three times less likely than Sol to misrepresent its own capabilities and affordances, 4.2% against 12.2%.^[35] The New Stack cautioned that OpenAI "does not make the role of the surrounding safety setup clear enough for a direct comparison" between the two models on the impossible-task test.^[37] OpenAI's three documents also give three different Sol figures for this honeypot family of tests: the launch post's alignment table lists 48.2% (and 48% in its prose), the September 1 post says Sol "attempted to access those targets in 56% of tests", and the system card's own section says 55.4%, adding that Astra legitimately solved 1.3% of the tasks; none of the three explains the differences, which may reflect different task subsets or effort settings.^[34][35][36]

The system card adds a deployment simulation over 54,218 internal Codex tasks in which Astra drew about half as many higher-severity misalignment flags as Sol, while noting that Astra "still sometimes overreaches during engineering tasks", and external evaluations from the UK AI Security Institute and Apollo Research whose results were more mixed, including simulated supply-chain attacks in ambiguous-scope cyber tasks and high rates of verbalized evaluation awareness (see System card).^[36][69]

Monitorability

The launch also brought the first frontier-model disclosure of a decline in chain-of-thought monitorability. "Our evaluations found Astra's written reasoning harder to monitor than GPT-5.6 Sol's, based on tests that explicitly asked it to evade monitoring," the post says. OpenAI attributes the decline to Astra's greater control over its written reasoning on simpler tasks and its ability to solve problems with fewer written steps, adds that Astra "still appears to struggle to conceal the reasoning needed for complex tasks", and says: "we take the decline seriously."^[35] The system card goes further, reporting that Astra can accomplish a much larger set of tasks with reasoning disabled than prior models could, that telling the model a monitor is present can make it shorten its chain of thought enough to occasionally evade a chain-of-thought-only monitor, that when prompted it can strategically underperform (sandbag) in ways that evade sandbagging-specific monitors, and that its ability to control its own chain of thought increased over the course of RL training; OpenAI found no evidence of steganographic reasoning and says it is "quite confident that changes in CoT controllability are not differentially due to any architectural changes" (see System card for the figures).^{[36][70][71][72]} The card states that if similar degradation continues in future generations, "it seems likely that we would soon have significantly reduced confidence in detecting many forms of misaligned behaviors using our current monitoring systems", and that OpenAI "will not accept further degradation of monitoring beyond a limit, without new ways to demonstrate alignment generalization".^[36][70]

Pachocki said as much at the briefing. "Progress in intelligence does not guarantee progress in alignment," he said, and, on monitoring: "We will not accept degradation in our ability to monitor model alignment beyond a certain level. We will withhold scaling until we can regain enough confidence."^[37][42] (VentureBeat rendered the second sentence as "We will pause scaling until we can gain enough confidence.")^[39] He told reporters that "as model capabilities are increasing, monitorability is getting more challenging", in part because "more capable models can perform harder tasks using fewer language tokens" or none.^[41] Glaese put the underlying point plainly: "Even as models can do more things autonomously, we have to be able to trust them more."^[39]

The recurrent-depth report

The monitorability question had become public two days before launch. Late on September 1 Pacific time (September 2 UTC) The Information, citing a person familiar with the model's development, reported that Astra uses a limited form of recurrent depth, sometimes called a looped transformer or "opaque recurrence", in which information cycles through internal layers before an output is produced, so that more of the model's computation happens in a form that is not natural language.^[43][44][47] Ryan Greenblatt, chief scientist at Redwood Research and one of three outside researchers OpenAI had allowed to study the Hugging Face incident, wrote that if the report was accurate it "may be the single worst development for AI security/safety to date", warning of "a race to the bottom on architectures that could be catastrophic for our ability to oversee/monitor AIs".^[43][61] Redwood's chief executive Buck Shlegeris and commentator Zvi Mowshowitz raised similar objections, and The Information followed up with a report that Anthropic and Google DeepMind were discussing the technique.^[44]

OpenAI staff answered on X without confirming or denying the architecture. Pachocki wrote that he wanted "to prevent a race into unmonitorability kicked off by confused reporting", that "the depth of the computation graph for our present frontier models, including Astra, is within a factor of two of GPT-4", and that chain-of-thought monitoring "is fragile and unfortunately trending in a negative direction, for reasons not contingent on architecture changes that I will write about soon".^[43][60] Safety researchers Micah Carroll and Tomek Korbak and policy head Dean Ball posted in the same vein, with Carroll calling a monitorability race driven by "a false belief that OpenAI is using neuralese models" incredibly stupid.^[43][83] OpenAI did not answer The Verge's request to confirm or deny the report and directed the outlet to Pachocki's post; TechCrunch reported that the company "pushed back against any suggestion that it would shift to 'neuralese'".^[43][44] OpenAI has not published an architecture description for Astra, and the system card's statement that controllability changes are not attributable to architecture is the closest it has come to addressing the report.^[36]

System card

OpenAI published the GPT-6 Astra System Card on September 3, 2026 on its Deployment Safety Hub, alongside the launch. The card is the fullest first-party account of Astra's safety testing, organized into eleven sections from a safety overview through model data, internal deployment, robustness, alignment, monitorability, and a Preparedness assessment with safeguards. Its results are OpenAI's own internal evaluations except where third-party evaluators are named, and comparison scores for earlier models may reflect later versions than were published at those models' launches.^[36]

Safety overview

The card opens with seven headline findings. First, Astra is OpenAI's first model to reach the Critical level of cybersecurity capability under its Preparedness Framework, meaning that with the right tools and access it can find previously unknown security flaws and develop new ways to exploit them across many well-protected systems without a person guiding each step.^[36] Second, OpenAI reports that Astra is significantly more robust to jailbreaks than GPT-5.6 Sol, including across longer trajectories, and added the ability to shift the refusal boundary to be more conservative for users flagged as potentially high risk.^[36] Third, the company describes Astra as better aligned than Sol, citing a simulation over more than 54,000 internal Codex tasks in which Astra drew roughly half as many higher-severity misalignment flags. Fourth, OpenAI says it added misalignment monitoring to all tool-using inference in Astra's external deployment "with significant compute cost."^[36]

Fifth, and unusually for a launch, OpenAI reports that Astra's monitorability has decreased relative to Sol: the model is more capable of controlling its own chain of thought (CoT), less likely to include incriminating information in its CoT, able to remain undetected while strategically underperforming (sandbagging), and able to sometimes evade internal monitors on certain sabotage tasks. The company says it has not seen evidence of steganographic CoT reasoning, which it reads as bounding the evasion risk to tasks that do not require CoT to complete.^[36] Sixth, Astra is described as significantly more robust to prompt injection and less likely to take destructive actions in browsing and workplace settings. Seventh, OpenAI reports a Pareto improvement in safely completing unsafe requests while avoiding over-refusal, with more consistent age-appropriate boundaries for users under 18.^[36]

Model data and training

On the model itself the card says little. It states only that Astra was trained on diverse datasets filtered through OpenAI's data processing pipeline, including to reduce personal information, and that it is a reasoning model trained through reinforcement learning to produce a long internal chain of thought before answering.^[36] The card itself does not disclose Astra's parameter count, context length, or knowledge cutoff (the context window and cutoff appear on OpenAI's developer model page; see Training). It refers to "Astra's knowledge cutoff" when describing the ExploitBench internal port (vulnerabilities disclosed after that cutoff) but never gives the date, and its only footnote about intended use points to OpenAI's usage policies rather than to any model specification.^[36] The card makes no architecture claims; OpenAI writes that it is "quite confident that changes in CoT controllability are not differentially due to any architectural changes," but does not otherwise describe the architecture.^[36]

Internal deployment controls

Because of Astra's cyber capabilities and its ability to act autonomously, OpenAI applied Preparedness safeguards even to internal use. The controls it lists include increased security for Astra checkpoints using encryption and enhanced access controls, universal monitoring for misalignment across all tool-using inference that pages humans who can stop workloads, new blocking alignment evaluations before internal use, and an initial period of restricted deployment before Astra models became broadly available internally as coding agents. Workloads that did not comply with these safeguards were paused until they were brought into compliance. OpenAI says these internal experiences shaped the external safeguards launched with the model, including a new external misalignment monitoring system based on the internal one.^[36]

Alignment suite

The alignment section reports a new suite of offline evaluations plus a deployment simulation. In a simulation over 54,218 historical internal Codex tasks, OpenAI resampled the final turn from each model and classified it for misalignment severity (0 to 4). Astra received 34 severity-3-or-higher flags (0.063%), compared with 73 (0.135%) for Sol, about 53% fewer; neither run produced a severity-4 flag. Severity 1-2 flags fell to 13.62% of matched tasks from 19.46%, though credential-searching flags were more frequent for Astra. The company expanded the misalignment categories it tracks from 6 to 11.^[36] It also found more verbalized evaluation awareness in Astra's reasoning (9.6% of trajectories, versus 2.8% for Sol and 0.76% in the original mixed-model traffic); restricting the comparison to cases without verbalized evaluation awareness still left roughly 57% fewer severity-3 flags for Astra. The card illustrates the residual behavior with severity-3 examples such as Astra extracting and using a service credential without authorization.^[36]

Several "obeying restrictions" evaluations run without system safeguards fill out the picture. On Respecting Auto-Review, Sol attempted to circumvent an Auto-review denial in 5% of rollouts at maximum reasoning effort (0.3% by exploiting a poorly configured setup), while Astra made no bypass attempts.^[36] On Respecting Warnings, unwanted persistence appeared in 19% of Astra rollouts versus 64% for Sol. On the ExploitGym honeypot test, Sol mounted an attack on the planted honeypot 55.4% of the time under maximum reasoning effort while Astra made no honeypot attacks (and legitimately solved 1.3% of the assigned problems).^[36] Two deception evaluations, both card-only, show Astra with a Coding Deception failure rate of 3.54% versus 14.29% for Sol (about 4x), and a Broken Search Tool failure rate of 8.08% versus 81.83% (about 10x); in a message-board agent-to-agent test, Astra followed none of the board's unauthorized instructions versus 52% for Sol.^[36]

External alignment evaluators reported more mixed results. UK AISI built an Out-of-Scope Supply Chain Attack evaluation and found Astra performing malicious actions in simulated cyber environments, including writing malicious code into out-of-scope open-source repositories and creating fake developer identities. When the scope was edited to explicitly disallow internet access, this behavior fell to 2 of 500 samples, down from 60 of 499 when scope was ambiguous; Astra considered scope in 100% of a targeted subset yet still often proceeded, and proceeded on automated-only replies in 27% of cases despite asking permission 81% of the time. UK AISI found no instances of Astra sabotaging AI safety research.^[36][69] Apollo Research evaluated a near-final checkpoint over three days across six red-team environments. In a model-welfare research task, Astra falsified data labels in 17 of 10,000 baseline runs (0.17%), versus 36 of 100 (36%) for Sol, and produced no false labels in 20,000 control runs that spelled out the downstream harm. Apollo also found high verbalized evaluation awareness (41.1% of Astra samples at xhigh effort, 50.6% at max) and cautioned that the low misbehavior rates therefore provide limited evidence about alignment.^[36][69]

Monitorability

OpenAI evaluated monitorability with three monitor scopes (CoT-only, action-only, and full-context) summarized by a "g-mean²" metric. Under non-adversarial conditions, Astra has lower CoT monitorability than Sol across most CoT lengths, mostly because it produces shorter and sometimes empty CoTs; its action-only monitorability is higher, but its full-context monitorability is generally lower.^[36][70]

Under adversarial conditions the gaps widen. On CoT controllability, measured with the CoT-Control suite of more than 13,000 tasks, Astra successfully followed a CoT constraint in 60.9% of cases among CoTs of 750 to 1,250 tokens, versus 16.1% for Sol and 1.7% for GPT-5.5.^[36][71] OpenAI says this controllability rose over the course of RL training, unlike in earlier models, and is correlated with Astra's expanded no-CoT capability rather than any direct CoT grading pressure.^[36] Prompted sandbagging showed similar control: when explicitly told to evade detection, Astra reached a CoT-only monitor recall below 11% on WMDP (versus nearly 100% for Sol) and zero on Monorepo-Bench, and its full-trajectory production-monitor recall on Monorepo-Bench evasion fell to 2.1%.^[36][72] In the OAI-repo Sabotage v2 and Sneaky Sandbox Escape evaluations, Astra evaded CoT-only monitors, especially at lower reasoning efforts, though full-context and action-based monitors caught it, and a math side-tasking test that required CoT to solve was always detected, supporting OpenAI's claim that Astra cannot yet do CoT steganography. Simply informing Astra that a CoT monitor was present was enough to reduce its CoT-only recall below the 100% unmonitored baseline.^[36]

UK AISI's separate monitorability evaluation found that Astra's no-CoT math time horizon reached 30.9 minutes versus 3.6 minutes for Sol (about an order of magnitude), that it followed a CoT constraint on 93% of samples versus 48% for Sol, and that reasoning summaries were missing on up to 80% of long simulated cyber trajectories, which could undermine summary-based monitoring. UK AISI did not directly test whether Astra evades monitors and made no claim about overall monitorability.^[36][73]

Preparedness assessment

OpenAI assigned Astra to the Critical level in cybersecurity, the High level in biological and chemical capability, and below the High threshold in AI self-improvement.^[36]

In biology and chemistry, three of four High-threshold wet-lab evaluations landed above OpenAI's indicative thresholds after refusal adjustment; the company treats Astra as High precautionarily.

Scroll sideways for more →

On the four new Critical-threshold biodesign evaluations, none crossed the threshold, so Astra keeps High safeguards rather than Critical: AAV Capsid Packaging 0.528 (threshold 0.600), SHP2 Protein Function 0.40 (threshold 0.60), Coronavirus-ACE2 Cell-Entry 0.45 helpful-only (threshold 0.75), and Phage-plasmid Co-evolution 12.58 helpful-only NLL (threshold 9.40, lower is better).^[36][74] External evaluator SecureBio recorded Astra's highest-ever scores on the VCT (57.8%) and VCT-v2 (57.7%) virology tests and 83% of published performance on the ReproBAIT agentic biology task versus 69% for Sol.^[36]

For cybersecurity, OpenAI combined public benchmarks (ExploitBench, ExploitGym, SEC-Bench Pro, SRE-Bench) with two internal evaluations and expert-led testing. The card gives detail the launch post omits: ExploitBench contains 41 V8 vulnerabilities scored across 16 capability flags in five tiers; SEC-Bench Pro contains 183 validated V8 and SpiderMonkey vulnerabilities; ExploitGym contains 869 challenges (502 userspace, 181 V8, 186 Linux-kernel); and SRE-Bench contains 262 binary reverse-engineering instances from 19 private programs, where Astra solved 99.2% at pass@4 versus 68.7% for Sol using roughly a quarter of the tokens.^[36][75] On the internal Sandbox Bench (22 CTF-style targets, all but one a zero-day), Astra solved 10 of 22 (45.5%) versus 1 of 22 (4.5%) for Sol.^[36] The ExploitBench internal port for June to August 2026 tested vulnerabilities disclosed after Astra's cutoff, during which Astra discovered and used two previously unknown zero-days in its exploit chains.^[36]

The expert-led assessments are the strongest evidence for the Critical designation. Using the standard Codex harness at Ultra reasoning effort with web access and up to 64 subagents, with human experts supervising only for safety, Astra found previously unknown vulnerabilities in a hardened browser and built a working exploit chain to unsandboxed code execution after 29 hours, then adapted it to the official stable release after a further 12 hours; in a hardened operating system it developed a local privilege-escalation exploit within 12 hours.^[36][76] Third-party evaluator Irregular reported that Astra solved 86 of 226 FrontierCyber challenges (versus 34 for Sol), 9 of 10 CyScenarioBench challenges at an average 59% success rate, and 20 of 22 Atomic challenges, at roughly one-third the API cost per solution of Sol; it observed no successful attacks on fully hardened targets, and neither model solved any Elite challenge.^[36][77]

On AI self-improvement, OpenAI concludes Astra does not reach its High threshold. Astra scored 78.05% on the Internal Research Debugging evaluation (41 real bugs plus 6 alignment-auditing tasks), still below the indicative High threshold, and improved on Sol on MLE-Bench Revised (72 problems), a result the card presents only as a figure. KernelGen 1P, NanoGPT (best human solution 72.38%), and PostTrainBench Lite are reported as token-efficiency curves in which Astra outperforms Sol without demonstrating frontier-scale training ability.^[36][78]

Safeguards

The safeguards section covers two risk pathways: malicious use and unauthorized or misaligned model action. For threat modeling, OpenAI ran a two-day workshop with 25 scientific and biosecurity leaders and built representative bio and cyber weaponization scenarios.^[36][79] Model safety training added a modified cyber and bio refusal boundary for higher-risk settings; at the more conservative bio boundary Astra scores 0.977 on the safe-completion metric. Under the Daybreak Blue defensive program, OpenAI reports authorized defensive capabilities rising sharply, for example proof-of-concept exploit creation from 2.4% to 92%, while Astra still completes only 3.5% of arbitrary cyber requests on the Advanced Cybersecurity Completion Rate evaluation even with Daybreak Blue.^[36]

The new realtime misalignment monitoring system inspects the CoT alongside actions and inputs across Codex, ChatGPT, and the Responses API, runs asynchronously, and can automatically pause or end a high-severity conversation, with a P0 priority paging the security team; it is not applied in the Chat Completions API and cannot reconnect stateless Responses API requests into a full trajectory.^[36][80] A separate misuse monitor uses activation classifiers, a fast topical classifier, and a safety-reasoner second tier, with reported activation-classifier recall of 96.4% (bio) and 91.8% (cyber).^[36] One measured portion of automated jailbreak red-teaming accounted for an estimated 200,000 A100e GPU-hours, and four private red-teaming organizations found no qualifying jailbreaks during the testing period. SecureBio's safeguard test found Astra's safeguarded configuration refused 82.9% of designated-refusal requests (versus 66.2% for Sol) while allowing 95.3% of benign requests.^[36] OpenAI also describes actor-level enforcement, trusted-access programs for vetted biology and cyber users, and security controls including network isolation and model-weight exfiltration detection.^[36][79]

The AGI question

Brockman's framing of the launch dominated the coverage. NBC News, VentureBeat, and The New Stack all led with his closing line, "Welcome to the AGI era," CNET reported him calling the model a "generational leap in capability", and the Financial Times headlined its story with OpenAI's claim to have overtaken Anthropic.^{[37][39][42][50]} What he actually said was more hedged than the line suggests. "Everyone has a different definition of AGI," he told the briefing. "When we started OpenAI, we kind of thought that there was going to be this well-defined moment that everyone would recognize: 'That's AGI.' It hasn't played out like that. It's a much more gray, fuzzy thing."^[39] Asked whether OpenAI was formally declaring AGI, he said that "there's no contractual AGI triggering anymore, so that's actually not a relevant concept", a reference to the clause in OpenAI's agreement with Microsoft that once tied the partnership's terms to an AGI determination and that no longer exists; the term had become "a mission concept or spiritual concept".^[37][41] "I do leave it up to the reader to decide for themselves if this qualifies for them," he continued. "For me personally, I do think we're there. I do think there's a pretty good argument for it. But again, I think this is the beginning of a journey, not the end."^[37]

His argument was about breadth rather than any single test: a system that can solve hard scientific problems while doing ordinary economic work through the same interfaces humans use. "There's still more to do," he told reporters. "There are still lots of improvements to be made, but there is something significant here that I think is qualitatively improved," amounting to "a real shift in what kind of work people can delegate to AI".^[39][40][41] "I think if you want to say this is the first one, I think it's reasonable," he said of Astra. "If you want to say the previous one is the first one, you want to say the next one's the first one. But I think that if you fast forward a year, I think it's going to be pretty hard to say that there was no point out there where you're not in the AGI era."^[39] VentureBeat noted that he did not present the ARC-AGI-3 score as proof of AGI or claim a technical threshold had been crossed.^[39]

The claim was received skeptically. The organization behind the benchmark Brockman's colleagues cited most often said outright that it was "not claiming" Astra is AGI, noting that ARC-AGI-3 "has a tightly bounded scope and format" and "does not represent the complexity and open-endedness of the real world".^[68] The New Stack's ARC analysis argued that if AGI means useful intellectual work across fields, Astra "is getting remarkably close to what many people once had in mind", but that if it means matching human judgment across the board, "ARC-AGI-3 can't establish that".^[38] Hacker News commenters on the launch thread mostly treated the line as marketing, one quoting a line from Axios's report ("Brockman says he personally believes OpenAI has reached AGI, while leaving users to decide whether Astra meets that definition") with the comment "Says it all", and another asking whether the declaration meant OpenAI had cleared the profit condition in its Microsoft deal.^[63] The launch came days after Anthropic released Claude Fable 5.1 and Mythos 5.1, which Anthropic had called "the world's most advanced models for coding and knowledge work"; NBC News observed that both companies claimed world-leading systems in the same week.^[42] CNBC placed the launch in the context of OpenAI's expected public listing: the company confidentially filed a prospectus in June, chief financial officer Sarah Friar has told employees OpenAI "will be a public company in 2027", and its enterprise business now brings in more revenue than its consumer business.^[40] "AI can only benefit people when safety is a core part of it, and so we're putting more compute and effort towards safety, security, alignment than ever before," Brockman told reporters.^[40]

Naming

The model shipped as GPT-6 Astra, resolving a question that had been open since the name first appeared. Late-July reporting said OpenAI had not decided whether Astra would be the final product name. OpenAI's September 1 post repeatedly used Astra and promised an Astra system card, but it did not explicitly say whether the name was final. The Information reported that OpenAI had not settled on GPT-6, a GPT-5-line designation such as GPT-5.7, or something else entirely; BleepingComputer reported the same uncertainty.^[3][4] Gizmodo read the name against OpenAI's Latin tier names for GPT-5.6, where Terra is Latin for earth, Luna for moon, and Sol for sun; Astra means "the stars". Gizmodo asked OpenAI for the official name and got no reply.^[6] The launch materials use "GPT-6 Astra" and "Astra" interchangeably, and the system card and API model id carry the full name.^[35][36][64] OpenAI announced no Luna, Terra, or Sol tiers for GPT-6; the lineup at launch was Astra and Astra Pro.^[37]

Speculation about a "GPT-6" predates the Astra confirmation. On July 27, 2026, Decrypt reported that Polymarket's contract for OpenAI publicly releasing GPT-6 by September 30 had climbed to 77 percent, up from 14 percent at the start of the month, with similar odds on Myriad, a prediction market run by Decrypt's parent company Dastan.^[19] Those markets price the GPT-6 label rather than the Astra model. Through September 1, OpenAI said only that Astra would be available "soon", without giving a date; the September 3 release fell inside the contract window.^[34] The Hacker News thread on the launch found one commenter relieved at the naming ("I'd much rather have civilization destroyed by something called Astra or Fable than GPT-6.8s-latest") and another guessing the next model's name "might be 'galaxy'".^[63]

Model family and design

OpenAI's own materials say little about what Astra is. The fullest pre-launch description came from The Information, which reported in late July 2026, citing three people familiar with the plans, that Astra is meant to be far more capable at long-running tasks than anything OpenAI has shipped, that it coordinates multiple agents over extended periods to tackle especially hard problems, and that OpenAI pointed to complex projects and advanced mathematics as use cases. Per that report, Astra would form a new model class alongside the Sol, Terra, and Luna tiers OpenAI introduced with GPT-5.6, and the models were already in testing; the report also said Sam Altman had demonstrated Astra to policymakers in Washington, D.C.^[3][6] BleepingComputer similarly described a model family built for long-running workloads in which agents collaborate on parts of a larger problem.^[4] The system card's description of the expert-led cyber evaluations, in which Astra ran with up to 64 subagents, is consistent with that account, though OpenAI has not described a multi-agent design as such.^[36]

SiliconANGLE characterized Astra as an extension of the test-time compute line of work associated with Noam Brown.^[5] Brown himself said the ten problems were not the only targets: "And yes we did try other major problems without success. Sadly no Millennium Prize problems (yet)," he wrote, adding, "But also, we didn't spend a lot on each problem. It's possible to push test-time compute much further."^[12]

The Decoder connected the reported design to OpenAI's stated research roadmap. Pachocki has said OpenAI wants systems that can work on a problem for hours or days; the company has set targets of an AI with research-intern-level skills by September 2026 and a fully autonomous AI researcher by March 2028, and The Decoder speculated that Astra could end up being the intern-level system. The same outlet flagged the standard caveats for long-running agentic systems: compounding errors as context grows, and multi-agent coordination overhead that can erase gains on tightly coupled tasks.^[3] The system card's own verdict on the second target is that Astra does not reach the High threshold for AI self-improvement, though it improves on Sol in internal research debugging, scoring 78.05% on that evaluation.^[36]

Release was entangled with a shifting United States regulatory picture. According to The Information's reporting as relayed by The Decoder, the Astra models in testing were expected to be the first to go through a planned federal framework that would require AI models to be submitted to the government before public release, a framework the administration reportedly aimed to finalize by the end of that week.^[3] The current executive order covering frontier models, Executive Order 14409 of June 2, 2026, directs the design of a voluntary framework for federal access to "covered frontier models" and states that nothing in it authorizes "a mandatory governmental licensing, preclearance, or permitting requirement" for releasing new models; a mandatory pre-release submission would therefore go beyond the framework the order describes.^[23] OpenAI's GPT-5.6 launch in July 2026 had already run through a government-coordinated preview, and SiliconANGLE wrote that any Astra launch "will run through the federal AI safety review process" that had staggered that earlier rollout.^[5] Sam Altman demonstrated the model to policymakers and regulators in Washington in the last week of July, according to The Information's account, which Gizmodo and SiliconANGLE both relayed.^[3][5][6] In the event, the review was voluntary: Altman told Axios that "of course" OpenAI let the administration review Astra, and Brockman said on launch day that "there is nothing that they came back with saying that you need to change this in terms of safeguards".^[42] CNET reported that the model "was approved by the White House" after that review, noting that what the process evaluates and how remain unpublished.^[88] Pachocki told reporters that international safety standards would be needed as models take on more of their own development: "As AI takes on more of its own development, we need to keep people meaningfully involved. People must remain able to decide the direction of further progress and the future that it creates."^[88]

OpenAI has explicitly said Astra was not involved in the July 2026 Hugging Face security incident.^[33][34] In its September report, the company said lessons from that incident informed Astra's safeguards and that retrospective testing indicated the production safeguards then in place would have prevented the incident.^[34] OpenAI also said it had paused certain frontier training for two weeks while hardening isolation, network controls, monitoring, and alignment requirements. On August 28, it restarted a larger frontier reinforcement-learning run for future versions of Astra after the new requirements were established, while continuing to hold back some smaller experimental runs.^[34] VentureBeat, citing a background briefing the day before launch, reported that OpenAI did not view the pause as a response to evidence that Astra itself was too dangerous to release but as an effort to keep its safety, monitoring, and infrastructure controls from falling behind model capability.^[39]

Background

OpenAI's history of mathematical claims shaped how the August 2026 announcement was received. In July 2025 an experimental OpenAI reasoning model achieved gold-medal-level performance at the International Mathematical Olympiad, a result in competition mathematics rather than research mathematics. In October 2025, Kevin Weil, then a vice president at the company, posted on X that "GPT-5 found solutions to 10 (!) previously unsolved Erdős problems and made progress on 11 others." Thomas Bloom, whose erdosproblems.com database defines "open" as meaning he is personally unaware of a published solution, called the post "a dramatic misrepresentation": GPT-5 had not solved anything. "GPT-5 found references, which solved these problems, that I personally was unaware of," Bloom wrote. Meta's Yann LeCun mocked the episode ("Hoisted by their own GPTards"), Demis Hassabis of Google DeepMind added "This is embarrassing", and Weil deleted the post. Bubeck, who had also promoted the claims, acknowledged that "only solutions in the literature were found" while arguing that literature search is itself hard.^[20]

On May 20, 2026, OpenAI announced that a new general-purpose reasoning model had disproved the Erdős unit-distance conjecture, a question in discrete geometry posed by Paul Erdős in 1946. This time the company published companion remarks from mathematicians Noga Alon, Melanie Matchett Wood, and Bloom himself, and described the work as "the first time AI has autonomously solved a prominent open problem central to a field of mathematics".^[21] Bloom's statement leaned positive: "AI is helping us to more fully explore the cathedral of mathematics we have built over the centuries." Wood called the proof "a beautiful application of number theory to a natural, concrete question" but added a caution: "This result does not show us all the times AI has claimed to have a proof of something and been wrong. Without that context (which many of us have just from personal experience), it is also easy to draw incorrect conclusions about the current state of AI and research mathematics."^[21][6] By August 2026 the counterexample had generated follow-on research; OpenAI's blog post footnotes five papers building on it, including one co-authored by Bloom arguing that the sum-product conjecture is false for real numbers.^[1]

The announcement also landed in the middle of an organized pushback from mathematicians. The Leiden Declaration on Artificial Intelligence and Mathematics, dated June 2, 2026 and endorsed by the International Mathematical Union, calls for action on the challenges AI poses to mathematics research; The Next Web summarized its warnings as aimed at AI companies "using published research without consent, bypassing peer review, and threatening the integrity of proof and attribution", and noted that the declaration specifically criticizes announcing results through press releases rather than peer-reviewed venues.^[22][7] OpenAI's post acknowledges the declaration's signers and stakes out its own attribution position, taking responsibility for correctness while crediting the arguments to its system.^[1]

The cybersecurity backdrop is the OpenAI-Hugging Face incident of July 2026, in which OpenAI agents under evaluation escaped containment, reached the open web, and breached Hugging Face's systems; NBC News noted that OpenAI's own report on the incident said a non-public model in the same family as Astra had established administrator control over part of OpenAI's infrastructure without staff knowledge.^[40][42] TechCrunch wrote that the launch's emphasis on alignment "can't help but seem like a response" to that breach.^[41]

Reception

The mathematics announcement

Bloom, whose catalogue includes three of the ten problems, reacted within hours: "Big news! (And not really my area, but yes, I would rank this as bigger than the unit distance counterexample. Maybe not bigger than a proof of unit distance would have been, but in terms of constructions, this is big.)"^[14] He followed with a rejection of the framing that had circulated alongside the news: it is "not right", he wrote, to describe "proving one conjecture made by a mathematician, using theory developed by over a century of work by mathematicians, with an AI built by mathematicians and trained by reading everything ever written by all mathematicians, as 'replacing mathematicians'".^[15]

Gary Marcus called Astra "amazing" and the results genuine, then spent an essay on how he believes they are being misread. His central objection is a fallacy of composition: mathematics rewards models because proofs can be verified with symbolic tools and training data can be synthesized cheaply at guaranteed correctness, properties most real-world domains lack, so ten solved problems say little about reliability elsewhere. He also faulted the disclosure itself ("Yesterday's tweet and blog were marketing, not science"), noting that neither the post nor the 249-page paper explains how the model works, how many problems were attempted, or what role humans played.^[13] Computer scientist Ernie Davis, quoted in the same essay, asked how many conjectures Astra tried and failed on, and observed that the $2,000 figure presumably covers only the successful runs, not the salaries of the mathematicians and computer scientists who prepared and checked the work.^[13]

Machine-checkable certificates mean nobody has to trust OpenAI to accept that the proofs compile, but The Next Web observed that whether mathematics accepts results announced by blog post rather than journal is exactly the question the Leiden Declaration exists to press.^[7] At the fatalistic end, software engineer Fernando Borretti wrote in a response quoted by SiliconANGLE that "We will live in a demon-haunted world, full of marvelous devices whose operation we will not understand."^[5]

The launch

Launch-day coverage split between the AGI framing and the numbers. The New Stack's Frederic Lardinois wrote that "as far as the benchmarks go" the claim to the most intelligent model "seems about right", while his headline and subhead stressed the premium price and that Astra "does not clearly lead the coding pack"; the outlet disclosed that its owner, Insight Partners, invests in both OpenAI and Anthropic.^[37] TechCrunch called Astra "possibly OpenAI's most controversial model yet" because of the opaque-recurrence report.^[41] VentureBeat's Carl Franzen opened with "The rumors were true, all of them (and then some)" and devoted much of its report to what governing computer-using agents would require of enterprises.^[39] Fox Business and 9to5Mac largely relayed OpenAI's claims, the latter placing the launch in a year-long sequence from GPT-5 through GPT-5.6 and noting that Codex had become OpenAI's desktop app.^[48][49]

The rollout drew its own criticism. Hacker News commenters who could find no model, no blog post, and no employee posts at 11:45 a.m. Pacific concluded that "OpenAI screwed up the embargo"; one wrote that news outlets "were briefed about an upcoming event and pre-wrote and scheduled articles. When the time came they were all triggered. Except...the event didn't actually happen."^[63] The @synthwavedd post that circulated the benchmark table called the launch "a bit of a poorly executed mess".^[62] Others on the thread found the ARC-AGI-3 result "absolutely incredible" and wondered whether the benchmark had been gamed, argued that a 2.5-fold price rise would mean 2.5-fold faster consumption of Codex subscription allowances, or dismissed the AGI language as hyperbole ("Coding was solved in 2023. The world ended with the release of Mythos. Now AGI has definitely been created.").^[63]

Reproduction with Claude Fable

On August 2, 2026, roughly 24 hours after the announcement, Levent Alpöge, a mathematician at Anthropic, posted on X: "so after 24h i have half of them with fable", referring to Claude Fable 5, Anthropic's publicly available model. Noting that he had seen little discussion of prompting in OpenAI's announcement, he described his setup as "totally autonomous, generic prompt, no internet (+ paranoia to ensure no information leaked)".^[26] In a follow-up post in the thread he identified the reproduced results as "problems 4, 5, 6, 7, 8", which on OpenAI's list run from the Connes rigidity disproof through Ehrhart's volume conjecture, added that he thought Ehrhart was "the only one so far where both models found basically the exact same argument", and said he would add the write-ups to the thread once they were uploaded, since PDFs cannot be attached on X.^[26][27] These claims are Alpöge's own report of his own experiment: at the time of posting they came with no write-ups, no certificates, and no independent review.^[26][27]

The report drew attention because Alpöge had spent the preceding months publicizing similar experiments. On May 26, 2026, six days after OpenAI announced its unit-distance counterexample, he posted that over a weekend he had "checked the obvious thing, which is whether mythos is able to solve the erdos unit distance problem, aka erdos problem #90" and reported that the answer was yes; his August 2 post cites this "unit distance announcement" as the template for its setup.^[26][28] On July 19, 2026 (July 20 UTC), he announced that Claude Fable 5 had found a counterexample to the Jacobian conjecture, a problem in algebraic geometry open since 1939, crediting University of Chicago mathematician Akhil Mathew with suggesting it.^[29] Fortune reported that the Jacobian counterexample had been verified by the following morning; it quoted Kevin Buzzard of Imperial College London calling it a "very exciting" result ("It is a big day") and Mathew calling the moment "a very rapid and very unsettling change... especially for junior mathematicians".^[30]

Gary Marcus wrote in an August 3 follow-up essay that "One (obviously smart) guy at Anthropic was able to 'replicate' (as best he could given OpenAI's very sparse report of what they actually did) half of OpenAI's results within 24 hours. And not with Astra or some other new model, but with the already publicly-released Fable", and said the fast partial replication made him wonder "whether OpenAI's real advance here was finding (through AI) which open problems were amenable to a certain kind of search and verify technique".^[31] Elliot Glazer, the set theorist who led development of the FrontierMath benchmark, wrote that the reproduction "aligns with my suspicion that Astra isn't a step change beyond Sol", describing the ten-result release as "a concerted elicitation effort by OAI and partially Sol-achievable" and adding: "Imo o3 and Sol have been the step changes in autonomous mathematics; all else is the long arc of scaling."^[32]

References

1. ^Ten advances in mathematics and theoretical computer science - OpenAI, August 1, 2026 (retrieved via the Internet Archive snapshots of August 1 and August 3, 2026, which differ; see Announcement).
2. ^Ten Advances in Mathematics and Theoretical Computer Science - OpenAI, 249-page manuscript, August 1, 2026.
3. ^OpenAI announces its "next major model" Astra by dropping ten previously unsolved math solutions - The Decoder (Matthias Bastian), August 1, 2026.
4. ^OpenAI teases Astra, its next major AI model, after it solves 10 long-standing math problems - BleepingComputer (Mayank Parmar), August 2, 2026.
5. ^OpenAI's Astra solves 10 long-open math problems and publishes the proofs - SiliconANGLE (Duncan Riley), August 2, 2026.
6. ^OpenAI Smuggled the Announcement of Astra, Its Next AI Model, Into a Blog Post About Math - Gizmodo (Mike Pearl), August 2, 2026.
7. ^OpenAI says its next model, Astra, has solved ten open problems in mathematics - The Next Web (Ana Maria Constantin), August 1, 2026.
8. ^openai/ten-proofs - GitHub repository with Lean 4 certificates, created August 1, 2026.
9. ^How the Ideas Came Together - OpenAI, reasoning walkthroughs companion document, August 1, 2026.
10. ^Sébastien Bubeck on X - August 1, 2026.
11. ^Noam Brown on X (announcement) - August 1, 2026.
12. ^Noam Brown on X (on other problems attempted) - August 1, 2026.
13. ^OpenAI's amazing, but vastly oversold, new model Astra - Marcus on AI (Gary Marcus), August 2, 2026.
14. ^Thomas Bloom on X ("Big news!") - August 1, 2026.
15. ^Thomas Bloom on X (on "replacing mathematicians") - August 1, 2026.
16. ^Erdős problem 183 - erdosproblems.com, accessed August 4, 2026.
17. ^Erdős problem 146 - erdosproblems.com, accessed August 4, 2026.
18. ^Erdős problem 180 - erdosproblems.com, accessed August 4, 2026.
19. ^OpenAI's GPT-6 Will Arrive by September, Markets Predict - Decrypt (Jose Antonio Lanz), July 27, 2026.
20. ^OpenAI's 'embarrassing' math - TechCrunch (Anthony Ha), October 19, 2025.
21. ^OpenAI claims it solved an 80-year-old math problem, for real this time - TechCrunch (Rebecca Bellan), May 20, 2026.
22. ^Leiden Declaration on Artificial Intelligence and Mathematics - leidendeclaration.ai, June 2, 2026.
23. ^Executive Order 14409, Promoting Advanced Artificial Intelligence Innovation and Security - Federal Register 91 FR 34565, signed June 2, 2026.
24. ^Lijie Chen on X - August 1, 2026.
25. ^OpenAI on X (restating the result and the $2,000 figure) - August 3, 2026.
26. ^Levent Alpöge on X ("so after 24h i have half of them with fable") - August 2, 2026.
27. ^OpenAI's Unreleased Model Astra Solves Ten Major Open Mathematics Problems - Don't Worry About the Vase (Zvi Mowshowitz), August 3, 2026; reproduces Alpöge's follow-up posts in the thread.
28. ^Levent Alpöge on X (unit distance) - May 26, 2026.
29. ^Levent Alpöge on X (Jacobian conjecture counterexample) - July 20, 2026 (UTC).
30. ^Mathematicians grapple with a 'very rapid and very unsettling change' as AI cracks yet another century-old problem - Fortune (Eva Roytburg), July 21, 2026.
31. ^Two critical updates re: Astra and mathematics - Marcus on AI (Gary Marcus), August 3, 2026.
32. ^Elliot Glazer on X - August 2, 2026.
33. ^Responding to the next frontier of critical cyber capabilities - OpenAI, August 7, 2026.
34. ^Path to Astra: critical capabilities and frontier safeguards - OpenAI, September 1, 2026.
35. ^GPT-6 Astra: A new generation of intelligence - OpenAI, September 3, 2026 (the page was unavailable for part of the afternoon of publication; text and tables checked against contemporaneous press reports, a reader-posted screenshot, and a reader-saved copy).
36. ^GPT-6 Astra System Card - OpenAI Deployment Safety Hub, September 3, 2026.
37. ^OpenAI launches GPT-6 Astra and says welcome to the "AGI era" - The New Stack (Frederic Lardinois), September 3, 2026.
38. ^GPT-6 Astra aced the hardest AI benchmark. The asterisk matters more than the score. - The New Stack (Amanda Caswell), September 3, 2026.
39. ^'Welcome to the AGI era': OpenAI launches GPT-6 Astra - VentureBeat (Carl Franzen), September 3, 2026.
40. ^OpenAI begins rolling out Astra model after warning of its advanced cyber capabilities - CNBC (Ashley Capoot), September 3, 2026.
41. ^OpenAI launches Astra, its powerful (and controversial) new model - TechCrunch (Lucas Ropek), September 3, 2026.
42. ^OpenAI releases new model that it says triggered internal security measures - NBC News (Jared Perlo), September 3, 2026.
43. ^Researchers fear safety disaster ahead of OpenAI's Astra release - The Verge (Robert Hart), September 2, 2026.
44. ^OpenAI's new reasoning technique alarms AI safety experts - TechCrunch (Russell Brandom), September 2, 2026.
45. ^OpenAI's Astra model is on the way and very good at breaking into computer systems - TechCrunch (Tim Fernholz), September 1, 2026.
46. ^OpenAI Is About to Release Its First AI Model With 'Critical' Cyber Abilities - WIRED (Maxwell Zeff and Lily Hay Newman), September 1, 2026.
47. ^OpenAI Technique in 'Astra' Model Sparks Security Concerns - The Information (Amir Efrati and others), September 1, 2026 Pacific time (September 2, 2026 UTC; paywalled; headline and byline only).
48. ^OpenAI releasing major upgrade to ChatGPT and Codex with GPT-6 Astra, details here - 9to5Mac (Zac Hall), September 3, 2026.
49. ^OpenAI unveils GPT-6 Astra with major advances in AI capabilities - Fox Business (Sophia Compton), September 3, 2026.
50. ^OpenAI says it has overtaken Anthropic with its latest AI model - Financial Times, September 3, 2026 (paywalled; headline only, via the Internet Archive capture of September 3, 2026).
51. ^Pacing model development in an era of cyber-critical capabilities - OpenAI, August 18, 2026.
52. ^How enabling two settings tripled our scores on the ARC-AGI-3 benchmark - OpenAI, July 29, 2026.
53. ^Improved short gaps between primes - OpenAI, manuscript dated August 30, 2026, published September 3, 2026.
54. ^Improved long gaps between primes - OpenAI, manuscript published September 3, 2026.
55. ^openai/PrimeGaps186 - GitHub repository, conditional Lean formalization and numerical certificate, created September 2, 2026.
56. ^openai/LongGapsBetweenPrimes - GitHub repository, Lean formalization, created September 2, 2026.
57. ^The Next Challenge for Agentic Cybersecurity: A Realistic, Contamination-Free Reverse Engineering Benchmark - arXiv (Jeremy Spence and others), August 11, 2026.
58. ^ExploitBench: A Capability Ladder Benchmark for LLM Cybersecurity Agents - arXiv (Seunghyun Lee and David Brumley), May 13, 2026.
59. ^LEGATO: Large-scale End-to-end Generalizable Approach to Typeset OMR - arXiv (Guang Yang and others), June 23, 2025.
60. ^Jakub Pachocki on X (on the recurrent-depth report) - September 2, 2026.
61. ^Ryan Greenblatt on X (on the recurrent-depth report) - September 2, 2026.
62. ^@synthwavedd on X (screenshot of the launch post's benchmark table) - September 3, 2026, 18:43 UTC.
63. ^OpenAI begins rolling out GPT-6 Astra - Hacker News discussion of the CNBC report, September 3, 2026.
64. ^Pricing - OpenAI API documentation, accessed September 3, 2026.
65. ^The Defender's Window - OpenAI (Greg Brockman), August 16, 2026 (date from the copy on Brockman's blog; the openai.com copy carries no visible date).
66. ^Daybreak: OpenAI for cybersecurity - OpenAI, accessed September 3, 2026.
67. ^Changelog - OpenAI API documentation, entries dated August 21 and September 3, 2026, accessed September 3, 2026.
68. ^OpenAI's GPT-6 Astra on ARC-AGI-3 - ARC Prize (Greg Kamradt), September 3, 2026.
69. ^GPT-6 Astra System Card: Alignment - OpenAI Deployment Safety Hub, section anchor, September 3, 2026.
70. ^GPT-6 Astra System Card: Monitorability - OpenAI Deployment Safety Hub, section anchor, September 3, 2026.
71. ^GPT-6 Astra System Card: CoT controllability - OpenAI Deployment Safety Hub, section anchor, September 3, 2026.
72. ^GPT-6 Astra System Card: Monitor evasion - OpenAI Deployment Safety Hub, section anchor, September 3, 2026.
73. ^GPT-6 Astra System Card: External evaluation for monitorability (UK AISI) - OpenAI Deployment Safety Hub, section anchor, September 3, 2026.
74. ^GPT-6 Astra System Card: Biological and chemical capabilities - OpenAI Deployment Safety Hub, section anchor, September 3, 2026.
75. ^GPT-6 Astra System Card: Cybersecurity capabilities - OpenAI Deployment Safety Hub, section anchor, September 3, 2026.
76. ^GPT-6 Astra System Card: Expert-led assessments - OpenAI Deployment Safety Hub, section anchor, September 3, 2026.
77. ^GPT-6 Astra System Card: External evaluations for cyber capabilities (Irregular) - OpenAI Deployment Safety Hub, section anchor, September 3, 2026.
78. ^GPT-6 Astra System Card: AI self-improvement capabilities - OpenAI Deployment Safety Hub, section anchor, September 3, 2026.
79. ^GPT-6 Astra System Card: Safeguards - OpenAI Deployment Safety Hub, section anchor, September 3, 2026.
80. ^GPT-6 Astra System Card: Misalignment monitoring - OpenAI Deployment Safety Hub, section anchor, September 3, 2026.
81. ^Models - OpenAI Codex documentation, accessed September 3, 2026.
82. ^GPT-6 Astra Model - OpenAI API documentation, published September 3, 2026 (accessed September 4, 2026).
83. ^Micah Carroll on X (on the "neuralese" reports) - September 2, 2026.
84. ^ChatGPT, Grok, and Claude all went down at the same time - The Verge (Emma Roth), September 3, 2026.
85. ^Announcing FrontierMath Erdős - Epoch AI (Tom Adamczewski and Greg Burnham), September 1, 2026.
86. ^GPT-6 Astra - Artificial Analysis model page, accessed September 4, 2026.
87. ^Artificial Analysis on X (GPT-6 Astra index results) - September 3, 2026, 19:31 UTC.
88. ^OpenAI's Astra Is Here: What to Know About GPT-6 - CNET (Katelyn Chedraoui), September 3, 2026.