# Gemini 4 Argon ties for third place on an AI benchmark Google helped build > Source: > Published: 2026-10-15 13:21:33+00:00 *Gemini 4 Argon hasn't appeared on Epoch AI's benchmark index at all, because Google is still keeping it behind limited access. Google helped fund the scoring method its rivals are being ranked on instead.* Epoch AI's Capabilities Index, updated October 1, currently has Claude Opus 5.5 leading at 167.35. GPT-6 Astra sits just behind at 166.51, with Claude Sonnet 5.5 at 165.2 and Claude Fable 5.1 at 164.82 rounding out the top four. Gemini 4 Argon doesn't appear on the list at all. Google announced Argon on September 30 as its flagship reasoning model for coding, science and long-running agent tasks. But it has kept the model on limited access rather than putting it through the public evaluations the index draws on. So there's no Google model anywhere near the top of the board. The Epoch Capabilities Index, or ECI, isn't a single test. It's a statistical model, built on item response theory, the same family of math used to calibrate standardized tests, that stitches together scores from more than fifty separate benchmarks into one scale. Epoch AI anchors the scale with two fixed points: Claude 3.5 Sonnet sits at 130, GPT-5 at 150. Harder benchmarks count for more, so a model only climbs the index by clearing the evaluations that still stump everyone else. That's a deliberate design choice, and it's why Epoch AI can compare Argon against a model like GPT-5 even though the two were never tested on identical benchmark sets. Here's the part that makes this interesting rather than just another leaderboard screenshot. The ECI's underlying methodology, laid out in a paper called "A Rosetta Stone for AI Benchmarks," was funded in part by Google DeepMind's own AGI Safety and Alignment team. Google helped pay for the ruler. The ruler, so far, hasn't measured Google's newest flagship at all. There's a point worth sitting with here: the gap between 167.35 and 164.82 sits inside the confidence intervals Epoch AI publishes alongside the scores, which for the top models range roughly four points wide. When the leaders overlap that tightly, a two-point edge isn't a demonstrated ordering, it's noise the index hasn't resolved yet. If you're a founder picking an API based on a leaderboard screenshot, that distinction matters more than the ranking itself. The real story in the ECI isn't who's on top this week. It's that Anthropic and OpenAI have converged so tightly that the index built to separate them can't cleanly do it anymore, while Google's newest flagship isn't even in the running yet. [Google Restricts Its Most Powerful AI Model to Vetted Cyber Defenders First](https://startupfortune.com/google-restricts-its-most-powerful-ai-model-to-vetted-cyber-defenders-first/) Google launched Gemini 4 Argon on September 30 and gave unrestricted access first to vetted cybersecurity defenders through its Fairwind Program, not to paying customers. The model leads rivals on most disclosed benchmarks but trails Claude Opus 5.5 on coding-agent tasks, and Google says it's gating the release over the model's offensive cyber... - [Google restricts Gemini 4 Argon to cyber defenders](https://startupfortune.com/google-restricts-its-most-powerful-ai-model-to-vetted-cyber-defenders-first/) - [how to access Google's Gemini 4 Argon model](https://startupfortune.com/google-restricts-its-most-powerful-ai-model-to-vetted-cyber-defenders-first/) That convergence is itself news. A year ago, benchmark gaps between frontier labs ran in the double digits. Now the top four scored models, Opus 5.5, GPT-6 Astra, Sonnet 5.5 and Fable 5.1, sit inside a three-point band on a scale where Claude 3.5 Sonnet, a model from under two years ago, anchors at 130. The frontier hasn't stalled. It's just stopped having a clear leader, at least among the labs the index can currently see. Google's own benchmark claims for Argon told a different story at launch. The company touted wins against GPT-6 Astra on several internal evaluations. Vals AI's independent index currently ranks Argon first among 41 models it tracks, with a score of 68.90%, ahead of Sonnet 5.5's 67.04%, Opus 5.5's 66.97% and GPT-6 Astra's 63.13%. That's not a contradiction, it's just two different benchmarks measuring two different things: a model can lead one index while trailing on another built from a wider net of tests. There's also the rollout angle, which the ECI number doesn't capture at all. Google hasn't given Argon a general release. The company rolled it out through what it calls the Fairwind Program, limited initially to trusted cybersecurity partners who get early access to help test the model's defensive capabilities before broader availability. Argon can generate up to 1 million tokens in a single response. That's a sixteen-fold jump from the 64,000-token ceiling on earlier Gemini models, and it's the kind of spec that matters more to someone building a long-running coding agent than any single benchmark point. As of this week, Google still hasn't set a public release date, saying access will expand gradually to paid API customers and Google AI Ultra subscribers. So what does Argon's absence from the ECI actually cost Google here. Not much in the near term, since the model is still limited to Fairwind Program partners rather than available for the kind of broad evaluation the index runs on. But it does mean Google's own benchmark claims for Argon are, for now, uncorroborated by the one index that none of the major labs controls. Every vendor in this market publishes self-reported numbers that routinely get disputed within days. Third-party placement like this is worth more to a skeptical buyer than another slide deck of Google's own benchmarks, and it's a gap Google will need the index to close once Argon reaches general availability. **Also read:** [Banks that financed the AI data center boom now want out of that debt](https://startupfortune.com/banks-that-financed-the-ai-data-center-boom-now-want-out-of-that-debt/) • [ASML says chipmakers are already booking tools for 2029 and 2030](https://startupfortune.com/asml-says-chipmakers-are-already-booking-tools-for-2029-and-2030/) • [An Israeli AI startup says it cracked a 38 year old graph theory bound for $2,000](https://startupfortune.com/an-israeli-ai-startup-says-it-cracked-a-38-year-old-graph-theory-bound-for-2000/) *This article is posted in [AI News](https://startupfortune.com/category/ai/), check it out for more related stories.* [Gemini 4 Argon Posts the Lowest Hallucination Rate of Any Top AI Model](https://startupfortune.com/gemini-4-argon-posts-the-lowest-hallucination-rate-of-any-top-ai-model/) Gemini 4 Argon posted a 15% hallucination rate on Artificial Analysis's AA-Omniscience benchmark, the lowest of any top-scoring model, fueling Reddit claims that hallucinations are solved. But the same benchmark puts Argon's accuracy at just 50%, meaning the gain is mostly about admitting uncertainty rather than being more correct. - [gemini 4 argon hallucination rate benchmark results](https://startupfortune.com/gemini-4-argon-posts-the-lowest-hallucination-rate-of-any-top-ai-model/) - [does gemini 4 argon actually solve hallucination problem](https://startupfortune.com/gemini-4-argon-posts-the-lowest-hallucination-rate-of-any-top-ai-model/) ## Join the discussion [Open in the community →](https://startupfortune.com/community/) Almost there. Sign in and your reply posts straight away.