Google's Argon Tops One Benchmark League and Trails Another

Google released Gemini 4 Argon to developers, enterprises and consumers worldwide on 13 October 2026, thirteen days after unveiling the model to a restricted group of cyber defenders. The launch post by Koray Kavukcuoglu, senior vice president at Google DeepMind and chief AI architect at Google, sets an introductory price of $2 per million input tokens and $10 per million output tokens, with cached input at 95 per cent off. A footnote to the same post says the standard price of $4 and $20 applies once the introductory period expires on 30 November 2026.

Access runs through the Gemini API in AI Studio and the Gemini platform, through Google Antigravity for Google AI Ultra subscribers, and through the Gemini app and AI Mode in Search for Ultra and Pro subscribers. Reuters reported the launch on 13 October. Google also said it is expanding its Fairwind Program, which hands the model to vetted security teams without cyber guardrails, to hundreds of additional partners.

Two leaderboards, two answers

Google's post claims Argon ranks first on the Vals Index, which weights professional tasks across finance, legal, coding and tax work by economic sector. VentureBeat reported on 13 October that Vals AI scored Argon at 68.90 per cent as of 12 October, against 67.04 per cent for Claude Sonnet 5.5 and 66.97 per cent for Claude Opus 5.5.

Artificial Analysis, which aggregates reasoning, knowledge, mathematics, coding and agentic tasks into a single Intelligence Index, puts Argon at 53. On the figures VentureBeat published on 13 October, that leaves it level with OpenAI's GPT-6 Astra and Claude Fable 5.1, and behind Claude Opus 5.5 at 58, Claude Sonnet 5.5 at 56 and Meta's Muse Wonder Full at 56. Both organisations test independently of Google.

The comparisons Google itself published when it first announced Argon on 30 September are a separate category of evidence. In those vendor-run tests, Argon scored 19.6 per cent on Harvey's Legal Agent Benchmark against 5.4 per cent for GPT-6 Astra and 3.8 per cent for Claude Opus 5.5, and 51.3 per cent on Zapier's AutomationBench against 42.5 and 41.4 per cent. The same set showed Argon losing: on Terminal-Bench 4.0, a command-line coding evaluation, Google reported Opus 5.5 at 66.4 per cent and Argon at 57.4 per cent.

As Parallax Nexus reported on 12 October, the 30 September release sent Argon first to Fairwind cyber defenders without guardrails while developers, enterprises and consumers waited with no date attached. That wait turned out to be under a fortnight.

What Google says the model did inside Google

The launch post is unusually specific about internal use. Google says Argon helped the Chrome team separate core webpage handling from a roughly 6 million line legacy folder of browser features inside Chromium's 40-million-line codebase, finishing an architectural refactor across hundreds of files in weeks rather than months. On the open-source Agent Development Kit, Google says Argon worked overnight from a one-paragraph goal and produced optimisations, since merged, that cut in-memory session preparation time by 40 per cent in long sessions and sped cold imports by 21 per cent. Inference throughput for the Gemma 4 E2B model rose from about 37,900 to more than 108,000 tokens per second on an Nvidia A100. Across Google engineering, the post claims agentic code submissions rose by more than 35 per cent while rollbacks fell by nearly a third. Every one of those figures is company-reported.

The safety numbers carry the same caveat, though Google says third-party partners took part in the measurement. The post puts attack success on the Gray Swan indirect prompt injection benchmark at 0.7 per cent across 15 adaptive attempts, unauthorised shortcut attempts at 2 per cent on coding evaluations, and external bio red teaming at under 1 per cent of the bio sub-tasks Google tests for.

The introductory price is the deadline that matters

The honest reading is that Google has bought six and a half weeks of attention rather than a settled lead. Buyers get until 30 November to test Argon at half what Claude Opus 5.5 lists for, and the per-token comparison flatters the model less than it looks, because cost per finished task depends on how many reasoning tokens a model spends getting there. Two things are not known. Google has not said how the advertised 1 million unified input-output token limit maps onto the largest response a developer can actually receive in one go, a point VentureBeat raised on 13 October and Google did not resolve in its materials. Nor has anyone outside Google replicated the Chromium or Gemma results.