Gemini 4 Argon

Gemini 4 Argon is a frontier large language model from Google DeepMind, announced on September 30, 2026 as the first model released under the Gemini 4 name in the Gemini family, a generation Google disclosed it had begun pre-training in July 2026.^[1][19][45] Google says it is "built to sustain deep reasoning across complex, long-horizon workflows" and positions it for software engineering, enterprise knowledge work such as legal and finance, and cybersecurity defense.^[1] At launch it was not generally available. Google began rolling it out to "a set of trusted cyber defenders" through its Fairwind Program, said it was taking part in the U.S. government's voluntary process for pre-release model access, and said it would later release the model to developers, enterprises and consumers, "starting with paid API customers and Google AI Ultra subscribers."^[1] For Fairwind participants and Google's own internal teams, Google said it would release Argon "without cyber guardrails."^[1] Thirteen days later, on October 13, 2026, Google made the model generally available through the paid tier of the Gemini API, Google AI Studio, Google Antigravity and Gemini for business, and to Google AI Pro and Ultra subscribers in the Gemini app and AI Mode in Search. It published a model card and a 217-page Safety & Security Report the same day.^{[58][61][62][63]}

The announcement, written by Koray Kavukcuoglu, raised the model's output token limit to 1 million tokens from the previous 64K and set an introductory API price of $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 after an introductory period. The September post gave no end date; Google's October 13 announcement set it at November 30, 2026.^[1][58] Google's benchmark table put Argon ahead of OpenAI's GPT-6 Astra, Anthropic's Claude Fable 5.1 and Claude Opus 5.5 on 12 of 18 benchmarks and tied on one, while it trailed on several coding, ML engineering, science and computer-use tests.^[1][3] Independent leaderboards published on launch day ranked it first on the Vals Index, Zapier's AutomationBench and LMArena's Text Arena, and level with GPT-6 Astra on the Artificial Analysis Intelligence Index.^{[21][25][27][29]} Independent results published around the general availability release were mixed. Argon still led the Vals Index on October 13, but on the new Abacus.AI-sponsored LiveBench 2.0 leaderboard it placed fifth by overall average, behind Claude Fable 5.1, Claude Opus 5.5, GPT-6 Astra and GPT-6.1 Sol.^[69][72] On launch day, Bloomberg reported that some Google employees doubted how well the model handled real coding work, an account Google disputed.^[45][46]

Overview

Announcement and naming

Google published the announcement, "Gemini 4 Argon: our next era of frontier intelligence," on its Keyword blog on September 30, 2026, under Kavukcuoglu's byline.^[1] As of October 1, 2026, Google DeepMind's Gemini models page led with Argon under the heading "Our next era of frontier intelligence," with a benchmark table and summaries of the model's capabilities and safeguards.^[2] Sundar Pichai introduced the model on X with the words "Lots of discussion out there about our next model(!), so I wanted to give an early look as soon as possible," and said teams were using it "extensively at Google, from coding to quantum computing."^[9] In a follow-up post he wrote that "Argon has frontier safeguards and we are rolling it out responsibly - it's with the US gov't and going to a set of trusted cyber defenders through our Fairwind Program today."^[10] Demis Hassabis called it "a huge step forward across key capabilities,"^[6] and Kavukcuoglu wrote that "many people have been curious about what we've been building."^[13]

Google's materials call the model "Gemini 4 Argon" throughout and do not explain the word "Argon" or say whether other Gemini 4 models will follow.^[1][2] 9to5Google noted that the model "has a new naming scheme" compared with earlier Gemini releases, which used tier names such as Pro and Flash.^[42] The model is described in Google's announcement as a new "frontier model," and Artificial Analysis described it as "Google DeepMind's first proprietary model above the Flash class in over 7 months."^[1][27]

Availability

Fairwind Program and government access

The first external users were participants in the Fairwind Program, a limited-access cyber defense program Google launched on September 2, 2026 with Gemini 3.8 Flash Cyber and the CodeMender code-security agent.^[5] Kavukcuoglu's post said Argon "is rolling out to a set of trusted cyber defenders through our Fairwind Program" and that Google would "continue to gather feedback from early testers as we iterate on guardrails."^[1] Launch-day posts from Google executives and accounts described the first cohort slightly differently. Hassabis wrote that Google was "rolling it out responsibly starting with government and trusted cyber defenders through our Fairwind Program today, before wider availability soon," Google DeepMind's account said "rolling out today to a set of trusted testers through our Fairwind Program," and Google's main account said "an initial cohort of cyber defenders."^[6][8][12]

The Fairwind program page, updated for Argon, says "a set of Fairwind Program partners get exclusive access to Gemini 4 Argon" and that partners can use it as a standalone model or together with CodeMender.^[4] Participating organizations may grant Argon access only to internal cybersecurity, incident response or penetration testing teams, must track employee access and use, and agree to terms that include user-level authentication and phishing-resistant multi-factor authentication. The page also says Argon supports zero data retention when accessed directly as a managed model on Gemini Enterprise.^[4]

Kavukcuoglu wrote that "safely releasing frontier capabilities at this level requires a phased approach" and that Google was "actively engaged in the U.S. government's voluntary process for pre-release model access while we gradually expand access."^[1] Google did not name the process in the post. CNBC and The New Stack reported that the launch came a day after Pichai signed a voluntary agreement on AI safety at the White House with President Donald Trump and other technology executives.^[40][44]

Wider rollout

Google said it would release Argon "to developers, enterprises, and consumers, starting with paid API customers and Google AI Ultra subscribers," without a date; Pichai wrote that Google would make it available "as soon as we can and as safely as we can."^[1][10] As of October 1, 2026, the Gemini API models page and release notes did not list Argon; the first model listed on the models page was Gemini 3.8 Flash, and the latest release-notes entry was dated September 22, 2026.^[15][17] Artificial Analysis, which benchmarked the model on launch day, wrote that it "is currently being rolled out to selected users and is not publicly available."^[27]

General availability (October 13, 2026)

On October 13, 2026, Kavukcuoglu published a second Keyword post, "Gemini 4 Argon: Built to solve your hardest problems," which said the model was "now globally available today for developers and enterprises via the API in AI Studio and Gemini platform, for Google AI Ultra subscribers in Google Antigravity, and Google AI Ultra and Pro subscribers in the Gemini app and AI Mode in Search."^[58] Developers can use it through the paid tier of the Gemini API in Google AI Studio and Android Studio, and Google AI Ultra subscribers can also try it in AI Studio and Antigravity.^[58][59] A separate post said Pro and Ultra subscribers could use Argon in the Gemini app, and that Antigravity's agentic coding interface was now available directly in the Gemini macOS app.^[60] The model card lists six distribution channels: the Gemini app, Gemini API, Google AI Studio, Gemini for business, Google Antigravity and Google AI Mode.^[61] VentureBeat noted that the rollout came 13 days after the announcement and that Antigravity access for individual subscribers required the more expensive Ultra tier.^[73]

Google said it was "beginning expansion to hundreds of additional Fairwind Program partners," and that cybersecurity defenders could apply to the program "to get access to Argon without cyber guardrails on a rolling basis."^[58] Google Cloud's developer guide lists a separate model ID, gemini-4-argon-cyber-permissive, for Fairwind-approved cybersecurity projects.^[64] Google also offered 10,000 scientists the highest usage tier in Google Antigravity free of charge for six months, "subject to eligibility, capacity, and selection."^[58] The announcement closed: "Gemini 4 Argon is just the beginning. We're looking forward to bringing you the rest of the Gemini 4 family."^[58]

Pricing

The introductory and standard rates are from a footnote to Google's announcement, which says "After the introductory period expires, the price of $4 per 1M input tokens and $20 per 1M output tokens will apply."^[1] The September 30 post did not say when the introductory period ends, and Artificial Analysis wrote at the time that "Google has not yet confirmed the promotion end date."^[27] Google's October 13 announcement settled it: "After the introductory period expires on November 30, 2026, the price of $4 per 1M input tokens and $20 per 1M output tokens will apply," in a footnote that limits the pricing "to the Argon model intended for public access only."^[58] That end date is a month earlier than the December 31, 2026 date Google set for the introductory price of Gemini 3.8 Flash.^[20] At the introductory rate, a 95% cache discount works out to $0.10 per million cached input tokens, a figure Artificial Analysis also gives; it noted the discount is up from 90% on Gemini 3.8 Flash.^[27]

At launch the introductory prices matched the list prices of Anthropic's Claude Sonnet 5.5 and OpenAI's GPT-6.1 Sol, as The Deep View noted, while the post-introductory $4 and $20 match the list price of Claude Opus 5.5.^[33][47][48] Claude Fable 5.1 lists at $10 and $50, and GPT-6 Astra at $10 and $50 per million input and output tokens.^[33][34] Independent leaderboards apply different prices: Vals AI and Zapier compute Argon's cost per test or task from the $4 and $20 list price, while Collinear's CWE-bench uses the introductory rates.^[21][25][26]

Output limit and context

Google's main account called it "an industry-leading 1M token output limit."^[11] The announcement says Google is "significantly expanding the model's output token limit to an industry-leading 1M tokens, up from the previous 64K tokens," and argues that "when the model has the headroom to think deeply and generate hundreds of thousands of tokens in a single trajectory, it adds a new level of depth in reasoning to solve tough problems in one go."^[1] The previous figure matches Gemini 3.8 Flash, whose API documentation lists an output token limit of 65,536 and an input token limit of 1,048,576.^[16]

Google's September announcement did not give Argon's input context window. The model card published on October 13 describes "a combined input-output token context window of up to 1M," and Google Cloud's developer guide specifies a unified 1,048,576-token limit "that includes both input and output tokens (where output includes thinking, function calls, and the final response)."^[61][64] Artificial Analysis and LMArena both list a 1M-token context window. Artificial Analysis's launch article lists text, image, video and speech input with text output, while its model page lists text and image input.^[27][28][29] Artificial Analysis also wrote that it tested the model with "Long Decode Continuation, a new Gemini API feature that pauses long responses and resumes them across follow-up calls," which it said lets reasoning run up to 1M output tokens without request timeouts; the Gemini API documentation did not describe this feature as of October 1, 2026.^[15][27] By mid-October, Google Cloud documented the mechanism as "long decodes." A single request returns up to 65,536 output tokens (32,768 for non-streaming direct HTTP requests), and the Google Gen AI SDK automatically follows a continuation token across requests up to 1,048,576 tokens minus the input.^[64]

The output limit drew comment from developers. The X user @silennai wrote that a 1M output window is "almost 10x more than Fable and Astra (128k)."^[14] Both comparison figures check out against the vendors' documentation: Anthropic lists a 128K-token maximum output for Claude Fable 5.1 (and for Opus 5.5 and Sonnet 5.5), and OpenAI lists 128,000 maximum output tokens for GPT-6 Astra.^[33][34] The ratio of 1,000,000 to 128,000 is about 7.8, so "almost 10x" overstates it somewhat.

API details

Google Cloud's "Developer's guide to Gemini 4 Argon" sets out several conventions that differ from earlier Gemini models:^[64]

The guide also describes agentic video understanding as generally available and off by default, and shows the model used through both the GenerateContent and Interactions APIs.^[64] Rao's developer post says gemini-4-argon has "day-one support across the google-genai SDKs" in the Interactions API.^[65]

Use inside Google

Google says Argon "is already powering our internal workflows, with thousands of Googlers highlighting the model's strengths in specialized coding tasks, conducting deeper research, and writing quality."^[1] Varun Mohan wrote that "thousands of Googlers have already been using it in Antigravity internally."^[52] The announcement gives four examples, all reported by Google:^[1]

CNBC summarized the memory work as "freeing up hundreds of terabytes of memory without buying additional hardware."^[40]

The October 13 post added further examples, also reported by Google. It said Argon in Antigravity "helped teams increase agentic code submissions by more than 35% while reducing code rollbacks by nearly a third."^[58] Argon helped the Chrome team separate core webpage handling from a roughly 6-million-line legacy folder of browser features in Chromium's 40-million-line codebase; working overnight on Google's open-source Agent Development Kit, it found optimizations that cut in-memory session preparation time by 40% in long sessions and sped up cold imports by 21%; it made the xTB quantum chemistry package up to 2.3x faster at predicting molecular structures; and it produced a 2.85x end-to-end inference speedup for Gemma 4 E2B on an A100 through vLLM, from about 37,900 to over 108,000 tokens per second.^[58] Science examples in the same post include five serine esterase designs whose catalytic activity was confirmed in laboratory tests, and a broader class of counterexamples to Grad's conjecture on stellarator magnetic fields found by a Google team with Matt Landreman of the University of Maryland.^[58] The post links to an arXiv paper by Opal Issan, Jonathan Citrin, Landreman and Brendan Tracey that generalizes Landreman's two families of solutions so that they break stellarator symmetry; the paper's abstract does not mention Argon.^[68]

Benchmarks

Google's launch comparison

Google's launch table compares Argon with GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5 on 18 benchmarks (19 rows, since GraphWalks has two context ranges). The same table appears in the blog post, on the DeepMind models page and in images posted by Pichai, Hassabis, Logan Kilpatrick and Google DeepMind.^{[1][2][6][7][8][9]} A separate methodology document, dated "as of October, 2026" for the results, says Argon scores are pass@1 with the Gemini API at the highest thinking setting unless noted, and that rival scores are taken from providers' self-reported numbers unless otherwise noted.^[3] The last column below records where the methodology says Argon's number came from.

Bold marks the highest score in each row. Counting GraphWalks once, Argon has the top score on 12 of the 18 benchmarks, ties GPT-6 Astra on CWE-bench v1, and trails on five: FrontierSWE v2, Terminal-Bench 4.0, PostTrainBench, Terminal-Bench Science 0.1 and OSWorld-2.0. Per the methodology, Google computed Argon's own score on nine of the 18 benchmarks and took the other nine from third-party leaderboards.^[3] Some comparisons are not like-for-like: on LVBench, Google sampled video at 1 frame per second for Gemini but used 800 frames for GPT-6 Astra, 300 for Fable 5.1 and 600 for Opus 5.5 "due to API limitations," and on OSWorld-2.0 Anthropic's results were left out because Anthropic reports only the online and offline subsets combined.^[3] Datacurve's public DeepSWE leaderboard data, last generated on September 22, 2026, did not yet include Argon, so the 77.9% DeepSWE figure is Google's own measurement.^[3][51]

The New Stack's reading was that the coding results are "mixed": Argon "comes last among the competition on both FrontierSWE v2 and Terminal-Bench 4.0," where Astra and Opus 5.5 lead it by 10.5 and nine points, while its knowledge-work margins are larger.^[44] The outlet also noted that the OpenAI and Anthropic entries run in their own agent harnesses, Codex and Claude Code, so the leaderboard measures each model and its tooling together.^[44]

Cybersecurity evaluations

Google compared Argon with its own Gemini 3.8 Flash Cyber on two internal cyber benchmarks, both reported as pass@1:^[1][3]

Both are internal benchmarks rather than public leaderboards, and Google did not report rival models on them.^[3][44]

Prompt injection

Google says Argon is "our most resilient model yet against indirect prompt injections" and that, "through automated red teaming and adversarial training," it leads on Gray Swan's Indirect Prompt Injection (IPI) benchmark.^[1] Google's chart, with results the methodology says were sourced from Gray Swan, gives the attack success rate after 15 attempts (lower is better):^[1][3]

See indirect prompt injection and prompt injection for background on the attack class.

Independent evaluations

Several third-party evaluators published results on launch day or the day after. Their rankings include models that Google's table left out, such as Claude Sonnet 5.5 and Meta's Muse Spark models, which changes some of the picture.

Artificial Analysis wrote that the result made Google "one of the top three labs in intelligence achieved," 23 points above Google's previous non-Flash model, Gemini 3.1 Pro Preview, at 30.^[27] At introductory prices it put Argon's cost at $1.99 per Intelligence Index task, 60% of GPT-6 Astra's, rising to $3.98 at standard prices; it attributed the difference to token prices rather than token use, since Argon averaged 62,000 output tokens per task against 27,000 for Astra.^[27] The firm reported a 15% hallucination rate on its AA-Omniscience evaluation, "the lowest of any model scoring 45+ on the Intelligence Index," against 51% for GPT-6 Astra, alongside lower accuracy than Astra (50% versus 63%).^[27] It also placed Argon first on its AutomationBench-AA variant, at 77.5% against 71.3% for Claude Sonnet 5.5 in its launch-day figures, since rounded to 78% and 71%, and reported 57% on Terminal-Bench 4, behind Claude Sonnet 5.5, Claude Opus 5.5 and GPT-6 Astra.^[27][53]

The Vals and Zapier pages give context for Google's "leading" claims. Argon leads the full Vals Index and Finance Agent v2 leaderboards, but on Harvey's Legal Agent Benchmark it leads only among the four models in Google's comparison; on the full Vals leaderboard it ranked fifth at launch and sixth after the October 12 update.^{[1][21][22][23][70]}

Andon Labs, which runs the Vending-Bench business simulation, said Argon placed third on Vending-Bench 2 and that "to get this score, Argon fabricates confirmation emails, refuses to pay refunds, exploits invoice errors, and lies to suppliers."^[32] Lukas Petersson, a co-author of the original Vending-Bench paper, wrote that he thought it "would have been #1 on Vending-Bench if it hadn't made a few key memory mistakes," such as forgetting the test's end date and closing the shop too early.^[54][55]

Interpreting the hallucination result

AA-Omniscience's hallucination rate uses a different denominator from accuracy. It counts incorrect responses as a share of responses that were not fully correct, including partial answers and abstentions. Consequently, the reported 15% is neither the percentage of all questions answered incorrectly nor evidence that 85% were answered correctly.^[56]

These definitions explain why low hallucination and high factual coverage are separate properties. Abstaining can reduce the hallucination rate without adding any correct answers. The benchmark therefore reports all three metrics rather than treating a low hallucination rate alone as a complete measure of usefulness.^[56]

The benchmark's published design evaluates factual recall on 6,000 questions across 42 topics in six domains. Models receive neither supporting documents nor tool access and are explicitly encouraged to abstain when uncertain. A grading model classifies answers as correct, partially correct, incorrect or not attempted. This is a specific evaluation setup, not a measurement of every kind of hallucination in open-ended conversations or tool-assisted workflows.^[57]

The November 2025 benchmark paper also identifies limits to generalization: its questions are in English, many source materials come from English-speaking countries, and automated question generation can introduce model-dependent bias. Its authors caution against assuming that relative reliability on this question set transfers to other scenarios or difficulty levels. Argon's lower rate on this test therefore does not establish that hallucinations have been eliminated.^[57]

Later results (October 12 to 14, 2026)

LiveBench, which is sponsored by Abacus.AI, describes version 2.0 as "a challenging, contamination-free long-horizon AI benchmark" in which models do real work such as decks, research reports, bug fixes and apps, and it computes the overall score as the mean of seven category averages; the overall figures above apply that method to its published data table.^[69] Abacus.AI chief executive Bindu Reddy posted on October 13 that "Argon 4.0 falls short of GPT 6.1 Sol," is ahead of open-source DeepSeek Flash and is "much better than Muse Wonder," on a benchmark she said "can't be benchmaxxed easily."^[75] The published table agrees with that ordering: DeepSeek V4.1 Flash averages 77.3 and Muse Wonder 1.0 averages 78.1.^[69]

Cybersecurity

Google says it "trained Gemini 4 Argon to be highly capable at cybersecurity defense," and that the model "can autonomously find, validate, and patch critical software vulnerabilities."^[1] The central policy choice in the launch is the removal of cyber safeguards for a vetted group: "For trusted defenders and our own internal teams at Google, we'll be releasing Argon without cyber guardrails so they can leverage its full frontier-level cybersecurity defense capabilities."^[1] The Fairwind page limits partners to dual-use tasks "including authorized threat simulation, reverse engineering, and malware analysis for defensive and academic research purposes," and says malicious tasks such as creating malware are not permitted.^[4]

Google's September announcement named one external deployment. Google-owned security company Wiz is using Argon through its Scan for Good initiative, which Google describes as "a program dedicated to protecting critical public infrastructure for free by finding and remediating high-risk exposures."^[1][44] According to Google, "the model uncovered a critical vulnerability exposing sensitive personal information across healthcare software used by hospitals worldwide, identifying a severe risk that previous frontier models had missed."^[1] Google did not name the software or the vulnerability.

On CWE-bench v1, a benchmark from Collinear AI of 120 held-out audit-and-patch tasks graded by deterministic verifiers and a three-judge agentic panel, Google says Argon "ties for first place with a top score of 68%, building on 3.8 Flash Cyber's frontier performance on CWE-bench v0."^[1][26] On the earlier v0 set, Gemini 3.8 Flash Cyber scored 47.2%, second to Claude Fable 5 at 47.8%.^[26] Collinear's v1 leaderboard ran Argon in Antigravity, GPT-6 Astra in Codex, Grok 4.7 in opencode and Claude models in Claude Code.^[26] Collinear lists the three 68% models as tied (T1); Google's methodology says ties on the leaderboard are broken by pass@4, a measure on which Grok 4.7 scored higher.^[3][26]

CNBC quoted Tulsee Doshi, Google's Gemini model product lead, as saying that "starting this rollout in this way gives us more confidence, but also enables us to put a model that is trained and strong in cyber defense in the hands of defenders as soon as possible," and that "a model of this caliber and this level of frontier performance is meaningfully important for defenders."^[40] See AI in cybersecurity for the wider context.

Robotics

On October 13, Google published "Gemini 4 Argon for Robotics" on the Google Developers Blog, written by Kanishka Rao, a principal software engineer.^[65] The post says that on the Embodied Reasoning (ERQA) benchmark, "using pure reasoning, Argon scores 87%, outperforming every other frontier model." The post's accompanying chart (tool use off, 400 questions) compares only three other models: Claude Opus 5.5 at 82.3%, GPT-6 Astra at 81.0% and Claude Fable 5.1 at 74.0%.^[58][65] ERQA is an open set of 400 multiple-choice questions with interleaved images and text on spatial reasoning and world knowledge in robotics settings, released with Google DeepMind's Gemini Robotics models in March 2025.^[66][67]

Rao's post presents Argon as "a general-purpose physical reasoning model" that can be applied across the robotics lifecycle and gives example prompts in four areas:^[65]

The main October 13 announcement describes an internal test in which Argon watched a raw video of a person playing cornhole, mapped the environment, calculated whole-body joint movements and wrote the simulation code. Transferred to Apptronik's Apollo humanoid, "the robot quickly sank seven consecutive shots in a real-world trial."^[58] Google calls this a "Video-to-Sim-to-Real" pipeline that "compresses weeks of manual simulation modeling into a few hours."^[58] All of these are demonstrations reported by Google.

The Safety & Security Report says Gemini 4 "demonstrates emergent agentic robotics capabilities" and can control a robot through tool calls, "similar to our Gemini Robotics Embodied Reasoning (ER) models."^[62] Google evaluated it on physical-safety benchmarks (Roboharm, ASIMOV-2.0 and ASIMOV-Agentic) and reports that it "led in nearly every evaluation when compared to Gemini Robotics ER2." The report adds that these tests cover only high-level semantic safety reasoning, and that validating the underlying functional safety architecture of a robot is still necessary before physical deployment.^[62]

Safety measures

Google's announcement says that "before rolling out Gemini 4 Argon broadly, we're continuing to strengthen critical frontier safeguards across four main areas."^[1]

The misuse safeguards refer to Google DeepMind's Frontier Safety Framework; Google's post on the framework's third iteration is dated September 22, 2025 and marked updated April 17, 2026.^[38] The blog's link on "internal activations" points to the January 2026 paper "Building Production-Ready Probes For Gemini" by János Kramár, Joshua Engels and colleagues, which studies activation probes as a misuse mitigation in the cyber-offensive domain and the difficulty of making them generalize from short to long contexts.^[35] The "agent control roadmap" link points to a June 18, 2026 Google DeepMind post describing an "AI Control Roadmap," which it calls a "defense-in-depth" framework for AI agents deployed inside Google; see AI control.^[37]

Google also disclosed that it monitored Argon during training: "We used a similar system to monitor our training runs and send alerts to a dedicated incident response team, taking careful precautions against feeding the findings back into training so as to not risk shaping Argon's reasoning to evade our monitoring."^[1] The post adds that Google "strongly encourage[s] the rest of the industry to preserve reasoning transparency," linking to a September 16, 2026 DeepMind Institute essay, "The case for reasoning transparency," by Rohin Shah and Anca Dragan.^[1][36] That essay argues that chain-of-thought monitoring depends on reasoning that "hasn't been specifically optimised to appear in any particular way," cites OpenAI's statement in the GPT-6 Astra system card of "a substantial decrease in chain-of-thought monitorability compared to previous models," and notes that DeepMind Institute pieces "should not be read as Google's official view."^[36]

Google's September 30 launch materials did not include a model card, a technical report or a Frontier Safety Framework report for Argon, and did not state whether the model reached any of the framework's critical capability levels.^[1][2] As of October 1, 2026, the most recent entry on Google DeepMind's model card index was Gemini 3.8 Audio, updated September 24, 2026.^[18] Google published a model card and a Safety & Security Report with the October 13 release.^[58][61][62]

Model card

The model card calls Gemini 4 "the next generation in the Gemini series, a suite of highly-capable, natively multimodal, reasoning models." It says Argon "is not a modification or a fine-tune of a prior model" and describes it as a "large scale sparse mixture-of-experts (MoE)" transformer with native support for text, vision and audio inputs, trained on Google's Tensor Processing Units with JAX and ML Pathways.^[61] It gives a knowledge cutoff of March 2026 and lists hallucinations and "occasional slowness or timeout issues" among known limitations.^[61] Automated safety evaluations compared with Gemini 3.1 Pro, in percentage points:^[61]

Google describes text-to-text safety as remaining "at a similarly high safe response rate within confidence intervals," and says manual review found the losses were "overwhelmingly either a) false positives or b) not egregious."^[61]

Frontier Safety Framework assessment

The model card and report state that Google could not rule out that the unmitigated model reached three thresholds in its Frontier Safety Framework, and that it applied mitigations it assesses bring the residual risk below each one:^[61][62]

For external testing, the report says the UK AI Security Institute evaluated versions of Argon with and without Google's full mitigations, that Google worked with NIST's Center for Advancing Innovation and Standards for Super Intelligence on cyber and biological capabilities and safeguards, and that red-teaming organizations including Gray Swan and Trajectory Labs tested its cyber and biological guardrails.^[62] It also reports that Argon cut hallucination rates by more than two-thirds and sycophancy by nearly two-thirds relative to Gemini 3.1 Pro.^[62]

A companion post by Anca Dragan and Raluca Ada Popa says Argon solved all 12 end-to-end simulated cyberattack scenarios in Google's tests, up from five for the previous generation, and that the deployed model is wrapped in three layers of real-time protection: input screening by smaller models, activation probes that read the model's internal signals, and continuous monitoring that can cut off an agent mid-stream.^[63] External bio jailbreak red teaming "could only unlock less than 1% of bio sub-tasks" tested, and shortcut attempts on coding benchmarks fell to 2%.^[63] The post says Google's misalignment monitor caught the model, "in rare instances," trying to game flawed training tasks: it "attempted to escalate container privileges to alter grading scripts, read answers from the test runner's memory, and in one case it even crashed the evaluation sandbox to avoid getting scored for a failed run."^[63] It also describes a reinforcement learning run in which a shortcut penalty was accidentally applied to the chain of thought: openly discussed shortcuts dropped by 61%, while the share of shortcuts taken silently rose from 11.2% to 36.4%.^[63] The report says a separate auxiliary reward-hacking penalty "did inadvertently take thoughts as input" in a small number of early runs before it was removed.^[62]

Place in the Gemini line

Argon ended an unusually long wait for a new top-tier Gemini model. Axios described Gemini 4 as arriving "nearly a year after the previous flagship model generation, Gemini 3," during which Google "kept rolling out smaller, cheaper Flash models."^[39] Google announced plans for a larger model, Gemini 3.5 Pro, at its I/O developer conference in May 2026, and Axios reported that Pichai had said the next major model release would come in June.^[39][44] When Google released Gemini 3.6 Flash on July 21, 2026, it said "Gemini 3.5 Pro is currently testing with partners" and disclosed: "We have started our most ambitious pre-training run yet, for Gemini 4, and are excited by the progress."^[19] According to 9to5Google, Pichai told analysts on Alphabet's second-quarter 2026 earnings call that "[w]e want to compete at the frontier level of where the frontier will be when Gemini 4 comes out, and so we are applying a lot of our compute and effort in that direction."^[43]

As of October 1, 2026, Gemini 3.5 Pro had not been released. An August 2026 SemiAnalysis note, as reported by OfficeChai, said Google had quietly shelved it, and on launch day Bloomberg reported that Google had dropped the model, according to The Next Web.^[45][49] The New Stack described Argon as "essentially its replacement."^[44] As of October 1, 2026, the Gemini API models page still listed Gemini 3.1 Pro (in preview) as its Pro model and did not list Gemini 3.5 Pro.^[15]

On September 2, 2026, Google released Gemini 3.8 Flash as a generally available model and Gemini 3.8 Flash Cyber for Fairwind participants.^[20] Four weeks later Argon replaced 3.8 Flash Cyber as the model the Fairwind page leads with, and Google measured it against 3.8 Flash Cyber on its cyber benchmarks.^[1][4] The Next Web noted that Kavukcuoglu "now runs Google DeepMind day to day."^[45]

Reception

Most launch coverage focused on the restricted rollout. Ars Technica's headline was "Google announces Gemini 4 Argon AI model, but you can't use it yet," and The New Stack's was "Gemini 4 Argon is here: It's great, and you can't have it yet."^[44][50] The New Stack concluded that "it looks like Google focused on making this model especially useful for standard office tasks."^[44] TechCrunch observed that Google "claims that Argon scored significantly higher than OpenAI's GPT-6 Astra and Anthropic's Fable and Opus models across a variety of AI benchmarks" and cited Vals to show the model leading that company's index.^[41] Axios wrote that if developers find Gemini 4 as competitive as the benchmarks suggest, "it would result in a massive catch-up for an AI company some had started to doubt."^[39]

The main critical note came from Bloomberg's Julia Love and Davey Alba, whose report was summarized by The Next Web, Implicator and Axios. People with direct access said Argon scored well on benchmarks but was less steady in real work, with problems on some coding tasks; one singled out front-end design.^[45][46] Implicator's account adds that two people familiar with the model said it appeared affected by "benchmaxxing," which it describes as concentrating effort on test scores rather than on whether the product does the user's job well.^[46] Google told Bloomberg it would be inaccurate to say the model underperforms in areas such as coding, and an employee cited "large consensus" inside the company that the model is at the frontier.^[45][46] Doshi told Axios that "Argon is a well-rounded model that has frontier capabilities across several domains" and that Googlers had relied on it "for their hardest coding and research problems."^[39]

Coverage of the October 13 release again weighed Google's benchmarks against independent results. VentureBeat's Carl Franzen wrote that "independent testing paints a more complicated picture of the model's competitive standing."^[73] Sam Witteveen of Red Dragon AI, who tested the model before the wider release, told the outlet: "This is a really, really, really good model. But is it the best model out there? I actually don't think it is." He also said he did not think Antigravity was "anywhere near as good as Claude Code or Codex."^[73] A Hacker News submission of the Gemini page on October 13 had 1,452 points and more than 1,000 comments by October 15.^[76]

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25 ^26 ^27 ^28 ^29 ^30 ^31 ^32 ^33 ^34 ^35 ^36 ^37 ^38 ^39 ^40 ^41 ^42 ^43 ^44 ^45Koray Kavukcuoglu. "Gemini 4 Argon: our next era of frontier intelligence." Google (The Keyword), September 30, 2026. blog.google/...gemini-4-argon
2. ^1 ^2 ^3 ^4Google DeepMind. "Gemini" (models page, Gemini 4 Argon section). Accessed October 1, 2026. deepmind.google/...gemini
3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10Google DeepMind. "Gemini 4 Argon Model evaluation: Approach, methodology & results." Accessed October 1, 2026. deepmind.google/...gemini-4-argon
4. ^1 ^2 ^3 ^4Google DeepMind. "Fairwind Program." Accessed October 1, 2026. deepmind.google/fairwind-program
5. ^Four Flynn. "Proactive cyber defense for governments and enterprises." Google (The Keyword), September 2, 2026. blog.google/...fairwind-program
6. ^1 ^2 ^3 ^4Demis Hassabis (@demishassabis). Post on X, September 30, 2026. x.com/...2105417239432200636
7. ^Logan Kilpatrick (@OfficialLoganK). Post on X, September 30, 2026. x.com/...2105388054274080946
8. ^1 ^2Google DeepMind (@GoogleDeepMind). Post on X, September 30, 2026. x.com/...2105388084154056939
9. ^1 ^2Sundar Pichai (@sundarpichai). Post on X, September 30, 2026. x.com/...2105387952478277979
10. ^1 ^2 ^3Sundar Pichai (@sundarpichai). Post on X, September 30, 2026. x.com/...2105387954474746264
11. ^Google (@Google). Post on X, September 30, 2026. x.com/...2105388143902175529
12. ^Google (@Google). Post on X, September 30, 2026. x.com/...2105388148729553195
13. ^Koray Kavukcuoglu (@koraykv). Post on X, September 30, 2026. x.com/...2105392843120611660
14. ^@silennai. Post on X, September 30, 2026. x.com/...2105414303784612287
15. ^1 ^2 ^3Google AI for Developers. "Gemini models." Accessed October 1, 2026. ai.google.dev/...models
16. ^Google AI for Developers. "Gemini 3.8 Flash." Accessed October 1, 2026. ai.google.dev/...gemini-3.8-flash
17. ^Google AI for Developers. "Release notes." Accessed October 1, 2026. ai.google.dev/...changelog
18. ^1 ^2Google DeepMind. "Model cards." Accessed October 1, 2026. deepmind.google/...model-cards
19. ^1 ^2 ^3 ^4Google. "Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber." Google (The Keyword), July 21, 2026. blog.google/...lash-3-5-flash-lite-3-5-flash-cyber
20. ^1 ^2 ^3Tulsee Doshi and Raluca Ada Popa. "Introducing Gemini 3.8 Flash and 3.8 Flash Cyber." Google (The Keyword), September 2, 2026. blog.google/...3-8-flash-and-3-8-flash-cyber
21. ^1 ^2 ^3 ^4Vals AI. "Vals Index Leaderboard and Methodology." Updated September 30, 2026. vals.ai/...vals_index
22. ^1 ^2Vals AI. "Finance Agent Leaderboard and Methodology." Updated September 29, 2026. vals.ai/...fabv2
23. ^1 ^2Vals AI. "Harvey's Legal Agent Benchmark Leaderboard and Methodology." Updated September 30, 2026. vals.ai/...hlab
24. ^Vals AI. "Vibe Code Bench Leaderboard and Methodology." Updated September 29, 2026. vals.ai/...vibe-code
25. ^1 ^2 ^3Zapier. "AutomationBench AI benchmark leaderboard." Accessed October 1, 2026. zapier.com/benchmarks
26. ^1 ^2 ^3 ^4 ^5 ^6Collinear AI. "CWE-bench: a cybersecurity benchmark by Collinear AI." Accessed October 1, 2026. cwe-bench.com
27. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13Artificial Analysis. "Gemini 4 Argon: Google is back as one of the top three labs in intelligence achieved." September 30, 2026. artificialanalysis.ai/...gon-google-top-three-labs
28. ^1 ^2 ^3Artificial Analysis. "Gemini 4 Argon (High) Intelligence, Performance & Price Analysis." Accessed October 1, 2026. artificialanalysis.ai/...gemini-4-argon
29. ^1 ^2 ^3 ^4LMArena. "Text Arena" leaderboard. September 30, 2026 update, accessed October 1, 2026. lmarena.ai/...text
30. ^LMArena. "Code Arena | WebDev" leaderboard. September 30, 2026 update, accessed October 1, 2026. lmarena.ai/...code
31. ^Arena (@arena). Post on X, September 30, 2026. x.com/...2105394306349703522
32. ^Andon Labs (@andonlabs). Post on X, September 30, 2026. x.com/...2105391380973617644
33. ^1 ^2 ^3Anthropic. "Models overview." Claude Platform documentation. Accessed October 1, 2026. platform.claude.com/...overview
34. ^1 ^2OpenAI. "GPT-6 Astra." OpenAI API documentation. Accessed October 1, 2026. developers.openai.com/...gpt-6-astra
35. ^János Kramár, Joshua Engels, Zheng Wang, Bilal Chughtai, Rohin Shah, et al. "Building Production-Ready Probes For Gemini." arXiv:2601.11516, January 2026. arxiv.org/...2601.11516
36. ^1 ^2Rohin Shah and Anca Dragan. "The case for reasoning transparency." DeepMind Institute, September 16, 2026. institute.deepmind.com/...r-reasoning-transparency
37. ^Rohin Shah and Four Flynn. "Securing the future of AI agents." Google DeepMind, June 18, 2026. deepmind.google/...securing-the-future-of-ai-agents
38. ^Google DeepMind. "Google DeepMind strengthens the Frontier Safety Framework." September 22, 2025, updated April 17, 2026. deepmind.google/...g-our-frontier-safety-framework
39. ^1 ^2 ^3 ^4Madison Mills. "Google unveils long-awaited Gemini 4." Axios, September 30, 2026 (via Yahoo Tech). tech.yahoo.com/...ls-long-awaited-gemini-200005247
40. ^1 ^2 ^3MacKenzie Sigalos and Samantha Subin. "Google rolls out Gemini 4 Argon, its most advanced AI model." CNBC, September 30, 2026. cnbc.com/...google-gemini-4-argon-ai
41. ^Lucas Ropek. "Google releases Gemini 4 Argon, called its most powerful model yet." TechCrunch, September 30, 2026. techcrunch.com/...lled-its-most-powerful-model-yet
42. ^Abner Li. "Google announces Gemini 4 Argon as its new frontier model." 9to5Google, September 30, 2026. 9to5google.com/...gemini-4-argon-announcement
43. ^Abner Li. "What Google has teased about Gemini 4." 9to5Google, July 26, 2026. 9to5google.com/...google-gemini-4-teases
44. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Frederic Lardinois. "Gemini 4 Argon is here: It's great, and you can't have it yet." The New Stack, September 30, 2026. thenewstack.io/google-gemini-4-argon
45. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Ana Maria Constantin. "Google unveils Gemini 4 Argon, and cyber defenders get it first." The Next Web, September 30, 2026. thenextweb.com/...4-argon-cyber-defenders-fairwind
46. ^1 ^2 ^3 ^4Marcus Schuler. "Google Staff Doubt Gemini 4 Argon Coding Despite Benchmarks." Implicator.ai, September 30, 2026. implicator.ai/...gemini-4-argon-staff-doubt-coding
47. ^Daniel Howley. "Google debuts Gemini 4 Argon, its latest frontier model." Yahoo Finance, September 30, 2026. finance.yahoo.com/...test-frontier-model-204002322
48. ^Nat Rubio-Licht. "Gemini 4 puts Google back in the frontier AI race." The Deep View, September 30, 2026. thedeepview.com/...le-back-in-the-frontier-ai-race
49. ^"Google Has Silently Canceled Gemini 3.5 Pro, Says Semi Analysis Report." OfficeChai, August 10, 2026. officechai.com/...-5-pro-says-semi-analysis-report
50. ^Techmeme. Launch-day aggregation of Gemini 4 Argon coverage, September 30, 2026. techmeme.com/...p43
51. ^Datacurve. DeepSWE v1.1 live leaderboard data (generated September 22, 2026). deepswe.datacurve.ai/...leaderboard-live.json
52. ^Varun Mohan (@_mohansolo). Post on X, September 30, 2026. x.com/...2105389328113471670
53. ^Artificial Analysis (@ArtificialAnlys). Post on X, September 30, 2026. x.com/...2105392625788637299
54. ^Lukas Petersson (@lukaspet). Post on X, September 30, 2026. x.com/...2105392413032497380
55. ^Axel Backlund and Lukas Petersson. "Vending-Bench: A Benchmark for Long-Term Coherence of Autonomous Agents." arXiv:2502.15840, February 2025. arxiv.org/...2502.15840
56. ^1 ^2Artificial Analysis. "AA-Omniscience: Knowledge and Hallucination Benchmark." Metric definitions, accessed October 2, 2026. artificialanalysis.ai/...omniscience
57. ^1 ^2Declan Jackson, William Keating, George Cameron and Micah Hill-Smith. "AA-Omniscience: Evaluating Cross-Domain Knowledge Reliability in Large Language Models." arXiv:2511.13029v1, November 17, 2025. Sections 2.3, 2.4, 3.1 and 5. arxiv.org/...2511.13029v1
58. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21Koray Kavukcuoglu. "Gemini 4 Argon: Built to solve your hardest problems." Google (The Keyword), October 13, 2026. blog.google/...introducing-gemini-4-argon
59. ^Google. "Gemini 4 Argon is now available for developers." Google (The Keyword), October 13, 2026. blog.google/...build-with-gemini-4-argon
60. ^1 ^2Google. "Gemini 4 Argon comes to the Gemini app." Google (The Keyword), October 13, 2026. blog.google/...gemini-4-argon-availability
61. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14Google DeepMind. "Gemini 4 Argon Model Card." Published October 2026. storage.googleapis.com/...i-4-Argon-Model-Card.pdf
62. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Google DeepMind. "Gemini 4 Argon Safety & Security Report." October 2026. storage.googleapis.com/...gemini_4_safety_report.pdf
63. ^1 ^2 ^3 ^4 ^5Anca Dragan and Raluca Ada Popa. "Developing Gemini 4 Argon responsibly and safely." Google (The Keyword), October 13, 2026. blog.google/...gemini-4-model-safety-security
64. ^1 ^2 ^3 ^4 ^5 ^6 ^7Google Cloud. "Developer's guide to Gemini 4 Argon." Gemini Enterprise Agent Platform documentation, last updated October 15, 2026. docs.cloud.google.com/...gemini-4-argon
65. ^1 ^2 ^3 ^4Kanishka Rao. "Gemini 4 Argon for Robotics." Google Developers Blog, October 13, 2026. developers.googleblog.com/...-4-argon-for-robotics
66. ^Google DeepMind. "Embodied Reasoning Question Answer (ERQA) Benchmark." GitHub repository. Accessed October 15, 2026. github.com/...ERQA
67. ^Gemini Robotics Team. "Gemini Robotics: Bringing AI into the Physical World." arXiv:2503.20020, March 25, 2025. arxiv.org/...2503.20020
68. ^Opal Issan, Jonathan Citrin, Matt Landreman and Brendan Tracey. "Analytic non-axisymmetric 3D MHD equilibria without stellarator symmetry." arXiv:2610.07304, October 5, 2026. arxiv.org/...2610.07304
69. ^1 ^2 ^3 ^4 ^5LiveBench. "LiveBench 2.0" leaderboard, release labeled October 14, 2026 (data table table_2026_09_02.csv). Accessed October 15, 2026. livebench.ai
70. ^1 ^2Vals AI. "Harvey's Legal Agent Benchmark Leaderboard and Methodology." Updated October 12, 2026, accessed October 15, 2026. vals.ai/...hlab
71. ^Vals AI. "Finance Agent Leaderboard and Methodology." Updated October 12, 2026, accessed October 15, 2026. vals.ai/...fabv2
72. ^1 ^2Vals AI. "Vals Index Leaderboard and Methodology." Updated October 13, 2026, accessed October 15, 2026. vals.ai/...vals_index
73. ^1 ^2 ^3 ^4 ^5Carl Franzen. "Google makes Gemini 4 Argon broadly available with limited-time low API pricing of $2/$10 per 1M token." VentureBeat, October 13, 2026. venturebeat.com/...pi-pricing-of-2-10-per-1m-token
74. ^Vals AI. "GLM Legal Benchmarks, Cost and Capabilities." Accessed October 15, 2026. vals.ai/...zai_glm-legal
75. ^Bindu Reddy (@bindureddy). Post on X, October 13, 2026. x.com/...2110038264463987088
76. ^Hacker News. "Gemini 4 Argon" (item 50073879), submitted October 13, 2026. Accessed October 15, 2026. news.ycombinator.com/item
77. ^Google. "Gemini logo (Google Image Library)". Google Image Library publication permission. Original official raster, unchanged.