Geminiexchange
DisputedResearch

Google launches Gemini 4 Argon, posting 77.9% on DeepSWE v1.1 and a 0.7% injection attack rate

Google released Gemini 4 Argon with a 1 million-token output limit and a 77.9% DeepSWE v1.1 score in its own comparison table, opening access to vetted cyber defenders first.

Why it matters

Argon reaches vetted cyber defenders before paid API users and runs without cyber guardrails, so external security teams get the model ahead of most customers. The cybersecurity advantage rests on Google's own comparison table, and the two published counts of it differ, which leaves the size of the lead unverified outside Google's testing. No token, protocol or trading venue is involved, so the direct effect on crypto markets is limited.

Where the sources disagree

How many benchmarks Gemini 4 Argon leads in Google's own comparison

12 of 18 benchmarks led

Decrypt

14 of 19 evaluations led or tied for first

Brave New Coin

Decrypt reports Argon leading 12 of 18 benchmarks, while Brave New Coin reports Google's published comparisons showing Argon leading or tying for the top result in 14 of 19 evaluations.

Whether Decrypt's reporting disclosed benchmark names and scores

No benchmark names or numerical scores beyond a relative lead

Bitcoin Insider

DeepSWE v1.1 named, 77.9% reported

Decrypt

Bitcoin Insider says Decrypt's report did not name the benchmarks or give scores beyond a relative lead, while Decrypt's own article names DeepSWE v1.1 and reports 77.9% alongside competing figures.

Key facts

Argon scored 77.9% on the DeepSWE v1.1 software-engineering test, against74.2% for Claude Opus 5.5, 74.1% for GPT-6 Astra and 67.4% for Claude Fable 5.1, while Gemini 3.6 Flash scored 49% on the same test in July.

77.9 %Disputed figureReported by Decrypt

Argon recorded a 0.7% attack success rate on Gray Swan's prompt-injection test, below the 1.0% posted by Claude Opus 5.5 and Claude Fable 5.1.

0.7 %Reported by Decrypt

Argon can produce up to 1 million tokens in a single reply, up from the 64,000-token limit of earlier Gemini generations.

1,000,000 tokensReported by Decrypt

The first external rollout is limited to vetted cyber defenders through the Fairwind Program with no cyber guardrails, ahead of paid API customers and Google AI Ultra subscribers.

Reported by Decrypt

Google calculated Argon's DeepSWE v1.1 score itself, while rival models' figures on that test came from a public leaderboard and company reports.

Reported by Decrypt

Google's own benchmark comparison is counted differently by the two outlets: Decrypt reports Argon leading 12 of 18 benchmarks, while Brave New Coin reports Argon leading or tying for first in 14 of 19 evaluations.

Disputed figureReported by Decrypt

Bitcoin Insider relayed a CryptoBriefing report that Google is facing internal employee skepticism about the Gemini model family, without stating whether the concerns apply to Gemini 4 Argon specifically.

Reported by Bitcoin Insider

What happened

Google DeepMind announced Gemini 4 Argon on 30 September as a frontier model for coding, office work and cyber defense, and Brave New Coin reports Google has been using it internally for coding, research, engineering and quantum computing. On the DeepSWE v1.1 software-engineering test, Argon scored 77.9%, ahead of Claude Opus 5.5 at 74.2%, GPT-6 Astra at 74.1% and Claude Fable 5.1 at 67.4%, while Gemini 3.6 Flash managed49% on the same test in July. Argon posted a 0.7% attack success rate on Gray Swan's prompt-injection test, below the 1.0% scored by both Claude Opus 5.5 and Claude Fable 5.1. The two crypto outlets covering the launch count Google's own comparison differently: Decrypt says Argon leads 12 of 18 benchmarks in Google's table, while Brave New Coin says Google's published comparisons show Argon leading or tying for the top result in 14 of 19 evaluations, and both note rivals score higher on several tests. Decrypt adds that Google computed Argon's DeepSWE result itself while rival figures came from a public leaderboard and company reports, and reports the rollout goes to vetted cyber defenders through the Fairwind Program without cyber guardrails before paid API customers and Google AI Ultra subscribers. Bitcoin Insider relayed two claims it did not detail or verify: a Forkast report that the launch closes a 'pricing triangle' of comparable pricing across OpenAI, Anthropic and Google with no figures given, and a CryptoBriefing report of internal Google employee skepticism about the Gemini family, which did not say whether it concerned Argon specifically.

How the story developed

  1. Official statement

    Google DeepMind announces Gemini 4 Argon to trusted testers

    Brave New Coin reports Google DeepMind announced Argon on 30 September as a model for coding, enterprise workflows and cybersecurity, with outside access limited to trusted cyber defenders through the Fairwind Program.

    Brave New Coin

  2. First report

    First report by Decrypt

    Gemini 4 Is Here, and Google’s Flagship Tops All Other AI Models on Cybersecurity

    Decrypt

  3. Update

    Bitcoin Insider relays Decrypt's report and CryptoBriefing's internal-skepticism story

    The aggregator restates Decrypt's cybersecurity claim without figures and reports, citing CryptoBriefing, that Google employees are skeptical about the Gemini family.

    Bitcoin Insider

  4. Update

    Bitcoin Insider relays Forkast's pricing-triangle framing

    Forkast is reported to see the release as closing a three-way convergence in flagship model pricing among OpenAI, Anthropic and Google, with benchmark results framed as the more significant part of the story and no figures given.

    Bitcoin Insider

  5. Independent corroboration

    Independently corroborated by Brave New Coin

    Google Launches Gemini 4 Argon, Challenging GPT-6 Astra With 1 Million Token Output Window

    Brave New Coin

  6. Update

    Brave New Coin publishes an independent account of the launch

    The report gives the 1 million-token output window, notes internal Google use of Argon, and counts Argon as leading or tying across 14 of 19 evaluations while conceding rivals score higher on several tests.

    Brave New Coin

How coverage built up

Independent sources over time, counted the way the consensus panel counts them: a republication of a wire story does not move the line.

30 Sept, 23:203 independent sources2 Oct, 17:31
  • 30 Sept 2026, 23:20 UTC: Decrypt — 1 independent source, 1 reports
  • 1 Oct 2026, 00:17 UTC: Bitcoin Insider — 2 independent sources, 2 reports
  • 2 Oct 2026, 17:31 UTC: Brave New Coin — 3 independent sources, 4 reports

Affected entities

Original sources