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]

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 whose end date Google did not give.^[1] 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]} Bloomberg reported on the same day 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]

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 post does not say when the introductory period ends, and Artificial Analysis wrote that "Google has not yet confirmed the promotion end date."^[27] By contrast, Google's announcement of Gemini 3.8 Flash gave a fixed end date of December 31, 2026 for that model's introductory price.^[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 announcement does not give Argon's input context window. 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]

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.

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]

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 ranks fifth.^{[1][21][22][23]}

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]

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 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.

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 launch materials do not include a model card, a technical report or a Frontier Safety Framework report for Argon, and they do 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]

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]

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