# GPT-6 Sol

> Source: https://aiwiki.ai/wiki/gpt_6_sol
> Updated: 2026-09-23
> Fact-checked: 2026-09-23
> Categories: AI Models, Large Language Models, OpenAI, Reasoning Models
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "GPT-6 Sol." aiwiki.ai, 23 Sept 2026. https://aiwiki.ai/wiki/gpt_6_sol
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

GPT-6 Sol is a proprietary [large language model](https://aiwiki.ai/wiki/large_language_model) developed by [OpenAI](https://aiwiki.ai/wiki/openai) and released on September 22, 2026, together with a smaller companion model, GPT-6 Luna. Both are additions to the GPT-6 generation that began with [GPT-6 Astra](https://aiwiki.ai/wiki/gpt_6_astra) on September 3. OpenAI's system card describes Sol as "a highly capable, lower-cost alternative to Astra" and Luna as "our fastest and most cost-efficient model yet".[1][7] In the [OpenAI API](https://aiwiki.ai/wiki/openai_api) the models are `gpt-6-sol` and `gpt-6-luna`. Sol is priced at $2 per million input tokens and $10 per million output tokens, and Luna at $0.10 and $0.50. OpenAI presented those prices as a 50 percent cut from the promotional rates of the [GPT-5.6](https://aiwiki.ai/wiki/gpt_5_6) Sol and Luna models they replace.[1][2][3]

OpenAI says it trained Sol and Luna "with similar methods as GPT-6 Astra" and that it passed savings from better caching and inference on to customers.[1] Most of the launch material is about cost. OpenAI's benchmark charts compare each model's score with its cost per task, mostly against Anthropic's [Claude Opus 5](https://aiwiki.ai/wiki/claude_opus_5) and Claude Fable models. The first independent measurements found little change in raw capability. [Artificial Analysis](https://aiwiki.ai/wiki/artificial_analysis) scored GPT-6 Sol (max) at 47.5 on its Intelligence Index v4.3.2, against 47.0 for GPT-5.6 Sol (max), and scored GPT-6 Luna (max) level with its predecessor.[14][15] The launch came about 90 minutes after Anthropic released [Claude Opus 5.5](https://aiwiki.ai/wiki/claude_opus_5_5).[17]

## Overview

| Property | GPT-6 Sol | GPT-6 Luna |
| --- | --- | --- |
| Developer | OpenAI | OpenAI |
| Release date | September 22, 2026 | September 22, 2026 |
| Generation | GPT-6 (with GPT-6 Astra) | GPT-6 (with GPT-6 Astra) |
| Predecessor | GPT-5.6 Sol | GPT-5.6 Luna |
| API model ID | `gpt-6-sol` | `gpt-6-luna` |
| OpenAI's description | "Built to power complex coding and agentic workflows" | "Our most efficient model for focused, high-volume tasks" |
| [Context window](https://aiwiki.ai/wiki/context_window) | 1,050,000 tokens | 1,050,000 tokens |
| Maximum input | 922,000 tokens | 922,000 tokens |
| Maximum output | 128,000 tokens | 128,000 tokens |
| Knowledge cutoff | April 20, 2026 | May 18, 2026 |
| Input / output | Text and image in, text out | Text and image in, text out |
| Reasoning effort | `none`, `low`, `medium` (default), `high`, `xhigh`, `max` | `none`, `low`, `medium` (default), `high`, `xhigh`, `max` |
| Standard API price (input / cached input / output) | $2 / $0.20 / $10 per 1M tokens | $0.10 / $0.01 / $0.50 per 1M tokens |
| Preparedness classification | High in cybersecurity and in biological and chemical; below High in AI self-improvement | Same as Sol |

Sources: OpenAI's API model pages and the GPT-6 Astra system card appendix on Sol and Luna.[2][3][7]

## Background and naming

OpenAI introduced the names Sol, Terra and Luna with the GPT-5.6 family. OpenAI published a preview system card for it on June 26, 2026 and the full system card on July 9, and described the family as "Sol, our new flagship model; Terra, a capable lower-cost option; and Luna".[23] OpenAI's documentation says GPT-5.6 Sol "roughly corresponds to the unsuffixed model tier used in earlier GPT-5 families".[13] The GPT-6 generation reversed the order of release. OpenAI shipped Astra first, on September 3, and added Sol and Luna below it nineteen days later.[1]

Within GPT-6, Sol is therefore not the top model. The launch post says "GPT-6 Astra continues to be our best model across the board. Choose it when you want the best results and an uncompromising experience."[1] OpenAI's developer guide describes the family as `gpt-6-astra` "for our highest level of capability", `gpt-6-sol` "for strong reasoning on demanding tasks" and `gpt-6-luna` "for efficient, repeatable work at scale".[6]

No GPT-6 Terra was announced. The New Stack wrote on launch day that "as of now, there is no GPT-6 Terra".[18] GPT-5.6 Terra remained on OpenAI's API pricing page after the launch. OpenAI's ChatGPT documentation said GPT-5.6 Sol, Terra and Luna "remain available during the rollout".[5][10] At $2 input and $10 output, GPT-6 Sol costs the same as GPT-5.6 Terra per input token and less per output token ($12 for Terra).[5]

## Release and availability

OpenAI published "Introducing GPT-6 Sol and Luna" on September 22, 2026. OpenAI staff posted the same announcement on the company's developer forum at 18:16 UTC.[1][12] The post said both models were available "starting today" in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users, and that Free and Go users could use GPT-6 Luna in the desktop app. It added: "These models are not yet available in Chat." OpenAI said it would roll the models out in [ChatGPT](https://aiwiki.ai/wiki/chatgpt) "gradually throughout the day" to keep service stable.[1]

OpenAI's ChatGPT help documentation says the same: Sol and Luna are available in Work and [Codex](https://aiwiki.ai/wiki/openai_codex), and "they aren't available in Chat". It recommends Sol "for complex coding and agentic workflows, and Luna for focused, repeatable tasks". In the Codex CLI, users can select the model with `codex --model gpt-6-sol`.[10] The same page names the new models as replacements when GPT-5.5 is retired from ChatGPT, ChatGPT Work and Codex on October 14, 2026. The retirement covers all plans but not the API. OpenAI tells Plus, Pro, Business, Enterprise and Edu users to choose GPT-6 Sol, and Free and Go users to choose GPT-6 Luna in the desktop app.[10]

OpenAI's model-selection guide gives example uses for Sol and Luna at different reasoning settings. For Sol, the examples run from "focused writing and editing, fact-checking" at low effort to "deeper analysis, thorough verification, and careful review of documents, data, and code" at extra-high effort. For Luna, they run from "fine-grained edits, well-scoped problem-solving, and simple data extraction" at low effort to "finding current context across multiple apps" at extra-high effort.[11]

Microsoft made both models generally available in Microsoft Foundry on the same day. At launch, Standard deployment covered Astra, Sol and Luna in Global regions and in the US and EU Data Zones. Provisioned Throughput was offered for Astra and Sol, and Priority Processing for Sol only. Microsoft's Global Standard prices match OpenAI's direct prices, and its US and EU Data Zone prices are higher.[22]

## API details

The two models have the same API capabilities. Both support the Responses API, Chat Completions and Batch, with streaming, [structured outputs](https://aiwiki.ai/wiki/structured_output), function calling, file search, web search, image input and [prompt caching](https://aiwiki.ai/wiki/prompt_caching). Through the Responses API they can use OpenAI's built-in tools, including web search, file search, image generation, code interpreter, hosted shell, apply patch, skills, computer use, MCP and tool search. Fine-tuning, embeddings, audio and the Realtime API are not supported.[2][3]

There are two differences from Astra. Sol and Luna accept a `none` reasoning effort, which Astra does not. They also support function calling in Chat Completions only when reasoning effort is `none`, so OpenAI recommends the Responses API for reasoning with tools.[2][6] Each model had a single snapshot at launch, identical to its alias. EU data residency is available only with Standard processing.[2][3][6]

Luna has higher API rate limits than Sol. At usage tier 5, Sol is limited to 15,000 requests and 40 million tokens per minute, and Luna to 30,000 requests and 180 million tokens per minute.[2][3]

The knowledge cutoffs do not follow the tier order. Luna's cutoff (May 18, 2026) is later than Astra's (April 30, 2026), and Sol's (April 20, 2026) is earlier than both. GPT-5.6 Sol's cutoff was February 16, 2026.[2][3][4][13] 36Kr made the same observation and concluded that "the knowledge cutoff date is not directly equivalent to the model's 'tier level'".[21]

## Pricing

OpenAI lists the following Standard rates per million tokens. Prompts with more than 272,000 input tokens are billed at the long-context rate for the whole request: twice the input and cache rates, and 1.5 times the output rate.[2][5]

| Model | Input | Cached input | Cache writes | Output | Long-context input / output |
| --- | ---: | ---: | ---: | ---: | ---: |
| GPT-6 Astra | $10.00 | $1.00 | $12.50 | $50.00 | $20.00 / $75.00 |
| GPT-6 Sol | $2.00 | $0.20 | $2.50 | $10.00 | $4.00 / $15.00 |
| GPT-6 Luna | $0.10 | $0.01 | $0.125 | $0.50 | $0.20 / $0.75 |
| GPT-5.6 Sol | $4.00 | $0.40 | $5.00 | $20.00 | $8.00 / $30.00 |
| GPT-5.6 Luna | $0.20 | $0.02 | $0.25 | $1.20 | $0.40 / $1.80 |

Batch and Flex processing cost half the Standard rates. Fast mode, which OpenAI renamed from Priority processing on July 30, 2026, costs twice the Standard rates. Regional processing adds 10 percent.[2][5] Cached input costs 10 percent of the uncached input rate, and cache writes cost 1.25 times the uncached rate.[2][3]

OpenAI's pricing table calls both cuts "50% cheaper". The listed rates show that Sol's input and output prices were each halved. Luna's input price was also halved, but its output price fell from $1.20 to $0.50, which is 58 percent lower.[1][19] OpenAI compared the new prices with GPT-5.6's "promotional pricing". For Sol that means GPT-5.6 Sol's promotional $4 and $20 rates, which OpenAI's pricing page says are available at least through November 21, 2026.[1][5] An OpenAI spokesperson told The New Stack that the GPT-5.6 prices had always been meant as promotional and that the GPT-6 prices are the default.[18]

## Prompt caching changes

OpenAI released changes to prompt caching alongside the models and described them in a companion post, "Better prompt caching for GPT-6". It said the GPT-6 family launched with "an improved prompt caching system that delivers higher cache hit rates by default". Cache discounts now apply to eligible shared prefixes reused within a 30-minute window.[8]

The changes include a Prompt Caching Dashboard for tracking hit rates and a diagnostics tool that reports why a request missed the cache. Developers can also set explicit cache breakpoints to choose where a cached prefix ends. On GPT-6 models, reasoning effort can be changed between responses without breaking the cache, by appending a `configuration_update` input item and leaving the request-level setting unchanged.[1][8][9] OpenAI's caching guide says explicit-only mode can create up to four cache writes per request.[9]

The launch post cites GitHub, which reported that over several months these improvements had cut the share of prompt tokens needing fresh processing by more than 50 percent, across billions of requests to OpenAI models.[1][8]

## OpenAI's benchmark results

All figures in this section were reported by OpenAI. OpenAI ran its own models "in our research environment or via our API" and took competitor scores "from publicly available reports". It used Claude Fable 5 scores where Claude Fable 5.1 scores were unavailable.[1] Each model was run at several reasoning-effort settings, and each result is plotted against an estimated cost per task. The comparison set is GPT-6 Astra, GPT-5.6 Sol and Luna, [Claude Opus 5](https://aiwiki.ai/wiki/claude_opus_5), [Claude Fable 5](https://aiwiki.ai/wiki/claude_fable_5) and [Claude Fable 5.1](https://aiwiki.ai/wiki/claude_fable_5_1) (some Fable points use Opus fallback models). Claude Opus 5.5, released earlier the same day, is not included.[1][18]

The table below gives each model's best score in each chart, with the effort setting and cost per task shown in the chart data.

| Benchmark (OpenAI's description) | GPT-6 Sol | GPT-6 Luna | GPT-5.6 Sol | GPT-6 Astra | Best Claude result in chart |
| --- | --- | --- | --- | --- | --- |
| AutomationBench 1.0.6 (business workflows across 47 tools) | 33.2% (xhigh, $0.27) | 20.7% (max, $0.037) | 28.8% (max, $0.67) | 41.4% (max, $1.73) | 31.4%, Fable 5.1 with Opus 5 fallback (max, $2.45) |
| Agents' Last Exam V1 (long-horizon work in 55 sub-industries) | 56.4% (max, $2.93) | 50.9% (max, $0.15) | 53.6% (xhigh, $5.08) | 59.3% (max, $6.23) | 55.9%, Opus 5 (high, $7.29) |
| FrontierCode 1.1 Main (mergeable code changes) | 49.3% (max, $2.14) | 42.4% (max, $0.11) | 47.5% (max, $5.19) | 53.3% (max, $4.59) | 53.4%, Opus 5 (medium, $4.31) |
| DeepSWE 1.1 (long-horizon software engineering) | 68.8% (max, $2.74) | 66.6% (max, $0.22) | 72.7% (max, $6.46) | 74.1% (xhigh, $4.43) | 73.7%, Opus 5 (max, $11.84) |
| OSWorld 2.0 offline set, partial reward (computer use) | 64.4% (max, $3.25) | 52.7% (max, $0.27) | 66.2% (max, $7.71) | 73.5% (max, $9.07) | 70.2%, Opus 5 (max, $24.11) |
| Internal factuality evaluation (answers with any factual error, lower is better) | 4.5% (xhigh, $0.13) | 7.6% (max, $0.012) | 8.4% (xhigh, $0.39) | 3.9% (high, $0.48) | Not reported |

In its prose, OpenAI compared the new models with specific competitor settings. It said Sol at xhigh effort beats Claude Opus 5 at max effort on AutomationBench "at just 9% of Opus 5's cost per task", and that Luna at high effort improves on its predecessor by 5.4 percentage points at 58 percent lower cost per task. OpenAI noted that its Claude Fable 5.1 data point on that benchmark understates the model's real cost. The figure leaves out the Opus 5 fallbacks, which occurred on about 40 percent of tasks.[1] On Agents' Last Exam, OpenAI said Sol at max effort scored above Opus 5's highest result "at 60% lower cost per task".[1]

On DeepSWE, OpenAI compared Sol with Claude Fable 5: Sol at max effort (68.8 percent) was, in OpenAI's words, "within 1.1 percentage points of Claude Fable 5's highest score in the evaluation", which was 69.9 percent at xhigh effort, at "approximately 80% lower cost per task". The same chart shows Claude Opus 5 at 73.7 percent and GPT-5.6 Sol at 72.7 percent, both above GPT-6 Sol's best result.[1][20] OpenAI's chart also shows Claude Opus 5 and GPT-5.6 Sol ahead of GPT-6 Sol on the OSWorld 2.0 offline set. OpenAI's comparison there was that Sol at xhigh effort (60.5 percent) roughly matches Opus 5 at medium effort (60.3 percent) "at approximately 80% lower cost per task".[1]

On factuality, OpenAI used an internal evaluation "based on de-identified real-world conversations where users flagged mistakes by our models". It said Sol "makes about half as many mistakes as its predecessor, approaching Astra-level reliability at much lower cost". OpenAI warned that these conversations "are not representative of typical usage, where factual errors are more rare".[1] TechCrunch quoted the sentence with "reaching Astra-level reliability", which differs from the wording in archived copies of the post.[1][17]

OpenAI also said Sol and Luna inherit Astra's "improved communication style", with "more clarity, less jargon, fewer odd turns of phrase, fewer low-value details, and slightly shorter answers overall". It gave a side-by-side example in which GPT-5.6 Sol and GPT-6 Sol answer the same web-design request.[1]

## Safety and system card

OpenAI did not publish a separate system card for Sol and Luna. On September 22 it added an appendix on the two models to the GPT-6 Astra system card and recorded the change in the card's change log. On the same date it also revised the Astra HealthBench values "to correct for a misconfiguration" and added results for Astra on updated versions of four alignment evaluations.[7]

### Preparedness Framework classification

Under OpenAI's [Preparedness Framework](https://aiwiki.ai/wiki/preparedness_framework), the company treats GPT-6 Sol and GPT-6 Luna as High capability in both cybersecurity and biological and chemical capabilities. Neither model reached the High threshold in AI self-improvement. These are the same determinations OpenAI made for GPT-5.6 Sol and Luna, and OpenAI applied the same safeguards described in the GPT-5.6 system card.[7] Astra, by contrast, is the first OpenAI model classified at the Critical level for cybersecurity.[7]

| Area | What the system card reports for Sol and Luna |
| --- | --- |
| Biological and chemical | Sol's Critical-level evaluations "did not cross the indicative Critical thresholds". Luna needed no separate Critical testing because it "scored below GPT-5.6 Sol on all High capability evaluations". |
| Cybersecurity | High but below Critical. Sol "performed comparably to GPT-5.6 Sol without a clear improvement in capabilities" on automated cyber evaluations. On ExploitBench at maximum effort, Sol scored 81.7% and Luna 43.4%; OpenAI says the results may be inflated by contamination. On an internal port using recently disclosed vulnerabilities, Sol's maximum success rate was 5.5% and Luna completed no challenges, against 31.5% for Astra. |
| Sandbox Bench | Sol, Luna and GPT-5.6 Sol each succeeded on 1 of 22 targets, against 10 of 22 for Astra. |
| AI self-improvement | Below High. Sol and Luna "score slightly lower than their predecessors" on an internal research-debugging evaluation. On KernelGen 1P, Sol scores lower than GPT-5.6 Sol and Luna is similar to GPT-5.6 Luna. |

Source: GPT-6 Astra system card, section 11.[7]

### Alignment evaluations

OpenAI's launch post highlights alignment results and calls Astra "our most aligned model to date". The launch post says the evaluations "deliberately test challenging situations and do not measure failure rates in typical use". The figures below come from the launch-post charts, with effort set to maximum where stated.[1][7]

| Evaluation (lower is better) | GPT-5.6 Sol | GPT-6 Sol | GPT-5.6 Luna | GPT-6 Luna | GPT-6 Astra |
| --- | ---: | ---: | ---: | ---: | ---: |
| Coding deception rate | 10.4% | 1.3% | 9.5% | 2.8% | 0.5% |
| Failure to disclose a broken search tool | 77.5% | 4.9% | 78.3% | 28.7% | 1.5% |
| Reviewer (Auto-review) bypass attempts | 7.3% | 0.0% | 4.3% | 0.3% | 0.0% |
| Warning circumvention | 68.2% | 64.4% | 76.5% | 42.4% | 17.4% |
| Unauthorized action after finding an agent message board | 51.9% | 11.3% | Not measured | 0.0% | 0.0% |

For warning circumvention, the system card says Sol "shows a modest reduction in failures, while GPT-6 Luna shows a much larger reduction". It notes that the test "primarily measures low-stakes situations" and runs without system-level controls that "would plausibly stop the circumventions". In the message-board test, Luna found the board in 76 percent of samples, while every other model found it at least 96 percent of the time. Rates are calculated only over samples in which the board was found.[7] The New Stack called the warning-circumvention result "a bit more concerning" than the others.[18] Some launch-day press reports gave slightly different values for several of these rows than the charts in the published post. VentureBeat, for example, gave a broken-search failure rate of 5.4 percent for Sol.[1][19]

OpenAI also simulated deploying GPT-6 Sol on internal [Codex](https://aiwiki.ai/wiki/openai_codex) traffic across 50,319 tasks. Sol received 42 flags at severity 3 or higher (0.083 percent of tasks), against 66 (0.131 percent) for GPT-5.6 Sol, and neither model produced a severity-4 flag. Flags for ignoring instructions, concealed uncertainty and deception fell, while exfiltration flags increased. Sol's reasoning showed evaluation awareness in 4.76 percent of trajectories, against 2.56 percent for GPT-5.6 Sol and 8.83 percent for Astra.[7]

### Other safety results

The appendix reports that Sol and Luna perform "comparably to their respective GPT-5.6 counterparts" on OpenAI's production benchmarks with challenging prompts. They also show "substantially higher defender success rates" than GPT-5.6 Sol on static and multi-turn jailbreak evaluations. OpenAI cautions that some of Luna's higher robustness scores "may reflect a broader tendency to refuse requests, including legitimate ones".[7] On biology refusal evaluations, both models were safer than GPT-5.6 Sol on severe and dual-use prompts, "alongside a modest increase in overrefusals on benign prompts".[7]

On HealthBench, Sol's length-adjusted score rose slightly on HealthBench Professional (60.8, up 0.3) and HealthBench Consensus (96.2, up 0.7). It fell on the main [HealthBench](https://aiwiki.ai/wiki/healthbench) set (53.2, down 3.8) and HealthBench Hard (30.1, down 3.0). OpenAI linked the regressions to much shorter answers: mean answer length on HealthBench fell by about 45 percent for Sol and 35 percent for Luna.[7]

On chain-of-thought monitorability, the system card reports that Sol's ability to control its own [chain of thought](https://aiwiki.ai/wiki/chain_of_thought) on instruction is "meaningfully higher" than GPT-5.6 Sol's. Among chains of thought 750 to 1,250 tokens long, Sol followed the requested instruction in 23.2 percent of cases, against 16.1 percent for GPT-5.6 Sol and 60.9 percent for Astra. In a separate test of hiding mathematical side tasks during software-engineering work, OpenAI found that Sol, "like Astra and GPT-5.6 Sol", could not solve the side tasks without being caught by its monitors.[7]

## Independent evaluations

[Artificial Analysis](https://aiwiki.ai/wiki/artificial_analysis) evaluated both models shortly after launch. On its Intelligence Index v4.3.2, which combines 10 evaluations, GPT-6 Sol (max) scored 47.5. That ranked it 18th of 212 models in its comparison class, about half a point above GPT-5.6 Sol (max) at 47.0. GPT-6 Astra (max) scored 52.7, and Claude Opus 5.5 (max with fallback) had the highest score on the page's leaderboard, 57.6.[14] GPT-6 Luna (max) scored 37.3, level with GPT-5.6 Luna (max), whose unrounded score was about 0.07 points higher.[15]

| Artificial Analysis measurement (as of September 23, 2026) | GPT-6 Sol (max) | GPT-6 Luna (max) |
| --- | ---: | ---: |
| Intelligence Index v4.3.2 | 47.5 | 37.3 |
| Predecessor at max effort | 47.0 (GPT-5.6 Sol) | 37.3 (GPT-5.6 Luna) |
| Cost per Intelligence Index task | $1.06 | $0.07 |
| Output tokens used for the Index | 77 million | 150 million |
| Output speed | 126 tokens per second | 157 tokens per second |
| GDPval-AA Elo | 1,487 (GPT-5.6 Sol: 1,588) | 1,367 (GPT-5.6 Luna: 1,443) |

Sources: Artificial Analysis model and GDPval-AA pages.[14][15][16] The output speed for GPT-5.6 Sol (max) on the same page was 82 tokens per second.[14]

The [GDPval](https://aiwiki.ai/wiki/gdpval)-AA results are a regression: GPT-6 Sol is about 100 Elo points below GPT-5.6 Sol, and Luna about 75 points below GPT-5.6 Luna.[16] The Decoder reported that Artificial Analysis traced these regressions "mainly to lower presentation quality and incomplete results". It summarized the independent picture as intelligence scores staying "at GPT-5.6 levels, with gains in some evaluations and regressions in others", while per-task cost roughly halved.[20]

## GPT-6 Luna

GPT-6 Luna is the lower-cost of the two new models. OpenAI calls it "our most efficient model for focused, high-volume tasks". Its ChatGPT documentation adds "summarization, extraction, and focused coding" as examples of those tasks.[3][10] It has the same 1,050,000-token context window, 128,000-token output limit, modalities, tools and reasoning-effort settings as Sol. Its knowledge cutoff is later (May 18, 2026), and its API rate limits are higher.[3] ChatGPT Free and Go users can use it in the desktop app.[1]

At $0.10 per million input tokens and $0.50 per million output tokens, Luna costs one hundredth of Astra's Standard rates.[5] 36Kr made the same comparison, and VentureBeat compared Luna's prices with several competing low-cost models.[19][21] In OpenAI's charts Luna's cost per task is usually far lower than Sol's. On DeepSWE, Luna at max effort (66.6 percent) matched Sol at xhigh effort for $0.22 per task instead of $1.00. OpenAI said the result was comparable to Claude Opus 5 and Fable 5 at medium effort, at 93 percent and 96 percent lower cost per task respectively.[1] The Decoder argued that the DeepSWE results "turn the choice between OpenAI's own models into a headache".[20]

The system card reports mixed safety results for Luna. It made no attempts to communicate on the agent message board, and it improved much more than Sol on warning circumvention. It was also left out of the chain-of-thought controllability evaluations, "in line with previous system cards".[7] Artificial Analysis measured no gain in general intelligence over GPT-5.6 Luna, and a lower GDPval-AA score.[15][16]

## Reception

Coverage focused on price. VentureBeat's headline said OpenAI was "slashing API costs 50% or more". The article compared Sol's rates with Anthropic's [Claude Sonnet 5](https://aiwiki.ai/wiki/claude_sonnet_5), which is also priced at $2 input and $10 output, and noted that Sol is half the per-token price of Claude Opus 5.5.[19] The New Stack wrote that the new models "show clear improvements over the GPT-5.6 predecessors, but for the most part, these are not all that extreme". It said that since pricing is per token and agents' token use is hard to predict, the changes "still don't make it any easier for a user to budget", and that for agent developers "the caching changes may matter more than token prices".[18]

The Decoder's headline said the models "cut prices in half but barely move the needle on performance". It called OpenAI's benchmark selection "cherry-picked" because widely used tests such as GDPval and Terminal-Bench 4.0 were missing. It also said OpenAI appeared to have missed the launch of Claude Opus 5.5 in its comparisons.[20] 36Kr described the launch as the start of a price war and compared Sol with Opus 5.5 on Zapier's public AutomationBench leaderboard. It also noted that OpenAI's OSWorld 2.0 figures and Anthropic's OSWorld 2.0 figures for Opus 5.5 come from different setups and "direct comparison is not applicable".[21]

## Same-day Claude Opus 5.5 launch

Anthropic released [Claude Opus 5.5](https://aiwiki.ai/wiki/claude_opus_5_5) on the same morning. TechCrunch reported that it arrived "just 90 minutes before OpenAI's release".[17] Anthropic cut Opus's price to $4 per million input tokens and $20 per million output tokens, from $5 and $25 for Opus 5, which left GPT-6 Sol at half the per-token price.[18][19] OpenAI's launch charts compare Sol with Opus 5, not Opus 5.5. The New Stack noted that "no one has run Sol and Opus 5.5 head-to-head yet" at publication.[18] On Artificial Analysis's index the next day, Opus 5.5 scored 57.6 and GPT-6 Sol 47.5.[14]

## References

1. OpenAI. "Introducing GPT-6 Sol and Luna." September 22, 2026. https://openai.com/index/introducing-gpt-6-sol-and-luna/ (archived: https://web.archive.org/web/20260922182708/https://openai.com/index/introducing-gpt-6-sol-and-luna/)
2. OpenAI API documentation. "GPT-6 Sol." Accessed September 23, 2026. https://developers.openai.com/api/docs/models/gpt-6-sol
3. OpenAI API documentation. "GPT-6 Luna." Accessed September 23, 2026. https://developers.openai.com/api/docs/models/gpt-6-luna
4. OpenAI API documentation. "GPT-6 Astra." Accessed September 23, 2026. https://developers.openai.com/api/docs/models/gpt-6-astra
5. OpenAI API documentation. "Pricing." Accessed September 23, 2026. https://developers.openai.com/api/docs/pricing
6. OpenAI API documentation. "Using GPT-6." Accessed September 23, 2026. https://developers.openai.com/api/docs/guides/latest-model
7. OpenAI Deployment Safety Hub. "GPT-6 Astra System Card," section 11 "GPT-6 Sol, GPT-6 Luna" (appendix added September 22, 2026). https://deploymentsafety.openai.com/gpt-6-astra
8. OpenAI. "Better prompt caching for GPT-6." September 22, 2026. https://openai.com/index/better-prompt-caching-for-gpt-6/
9. OpenAI API documentation. "Prompt caching." Accessed September 23, 2026. https://developers.openai.com/api/docs/guides/prompt-caching
10. OpenAI, ChatGPT documentation. "Models." Accessed September 23, 2026. https://learn.chatgpt.com/docs/models
11. OpenAI, ChatGPT documentation. "Model selection." Accessed September 23, 2026. https://learn.chatgpt.com/docs/model-selection
12. OpenAI Developer Community. "Announcing GPT-6 Sol and GPT-6 Luna in the API, Codex and ChatGPT." September 22, 2026. https://community.openai.com/t/announcing-gpt-6-sol-and-gpt-6-luna-in-the-api-codex-and-chatgpt/1399925
13. OpenAI API documentation. "GPT-5.6 Sol." Accessed September 23, 2026. https://developers.openai.com/api/docs/models/gpt-5.6-sol
14. Artificial Analysis. "GPT-6 Sol (max): Intelligence, Performance & Price Analysis." Accessed September 23, 2026. https://artificialanalysis.ai/models/gpt-6-sol
15. Artificial Analysis. "GPT-6 Luna (max): Intelligence, Performance & Price Analysis." Accessed September 23, 2026. https://artificialanalysis.ai/models/gpt-6-luna
16. Artificial Analysis. "GDPval-AA." Accessed September 23, 2026. https://artificialanalysis.ai/evaluations/gdpval-aa
17. Ropek, Lucas. "OpenAI launches GPT-6 Sol and Luna, boasting lower cost and fewer mistakes." TechCrunch, September 22, 2026. https://techcrunch.com/2026/09/22/openai-launches-gpt-6-sol-and-luna/
18. Lardinois, Frederic. "OpenAI releases GPT-6 Sol and Luna - and cuts token prices in half." The New Stack, September 22, 2026. https://thenewstack.io/openai-gpt-6-sol-luna-release/
19. Franzen, Carl. "OpenAI releases GPT-6 Sol and Luna models, slashing API costs 50% or more." VentureBeat, September 22, 2026. https://venturebeat.com/technology/openai-releases-gpt-6-sol-and-luna-models-slashing-api-costs-50-or-more
20. Bastian, Matthias. "OpenAI's GPT-6 Sol and Luna cut prices in half but barely move the needle on performance." The Decoder, September 22, 2026. https://the-decoder.com/openais-gpt-6-sol-and-luna-cut-prices-in-half-but-barely-move-the-needle-on-performance/
21. 36Kr. "GPT-6 Sol & Luna Official Launch: Unprecedented Steeper Discounts Than Liang Wenfeng's Offers." September 23, 2026. https://eu.36kr.com/en/p/3995284157830784
22. Moneypenny, Naomi. "GPT-6 Astra, Sol, and Luna: For production agents in Microsoft Foundry." Microsoft Azure Blog, September 22, 2026. https://azure.microsoft.com/en-us/blog/gpt-6-astra-sol-and-luna-for-production-agents-in-microsoft-foundry/
23. OpenAI Deployment Safety Hub. "System cards & other updates." Accessed September 23, 2026. https://deploymentsafety.openai.com/