# Google's Gemini 4 Argon Raises Output to One Million Tokens and Leads 12 of 18 Disclosed Benchmarks

Argon scores 77.9% on the DeepSWE v1.1 long-horizon coding suite but trails GPT-6 Astra by 10.5 points on the harder FrontierSWE v2, and general access is gated behind a vetted-defender program.

- Published: 2026-10-02T06:31:19.109Z
- Canonical: https://polylog.news/ai/2026-10-02/google-s-gemini-4-argon-raises-output-to-one-million-tokens
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Google DeepMind](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/), [Polylog editors](https://polylog.news)

Google has released [Gemini 4 Argon](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/), positioning it for long multi-step work: agentic software engineering, financial and legal analysis, enterprise workflows and cyber defense. The structural change is the output budget. Argon can emit up to one million tokens in a single generation, [up from a 64,000-token ceiling](https://t.me/ai_machinelearning_big_data/11055) on prior Gemini models, which removes the need to checkpoint and re-prompt through long refactors, migrations and document reviews.

The benchmark picture is mixed in a way the launch materials do not fully convey. Across 18 disclosed evaluations, Argon [leads outright on 12 and ties on one](https://venturebeat.com/technology/google-unveils-gemini-4-argon-retaking-benchmark-lead-over-openai-and-anthropic-but-in-limited-release). On DeepSWE v1.1 it scores 77.9% against 74.2% for Claude Opus 5.5 and 74.1% for GPT-6 Astra. On the harder FrontierSWE v2 it scores 55.0%, behind Astra at 65.5% and Opus 5.5 at 62.3%. All of these numbers come from Google's own disclosure, with no independent reproduction yet.

Introductory pricing is $2 per million input tokens and $10 per million output tokens, matching Anthropic's Sonnet 5.5 rate card. Access is restricted at launch to vetted cyber defenders through Google's Fairwind program, with paid API customers and Google AI Ultra subscribers to follow.

## What this means

The competitive axis here is which model can sustain autonomous work longest per API call. A million-token output budget mainly benefits workloads where a model must hold a plan across hours of tool calls, which is exactly the segment where Anthropic's Claude Code and OpenAI's Codex have retained developer usage. Google is exposed in distribution rather than capability: gating the model behind a defender-vetting program creates an appearance of safety caution but slows early adoption, and the FrontierSWE gap gives rivals a basis for a counterclaim. When Google shipped Gemini 2.5 Pro with a long context window in 2025, the practical constraint turned out to be retrieval quality and cost at depth rather than the window itself. The likelier outcome over the next few weeks is that independent harness results on long-horizon suites land below Google's published figures while still clearing Gemini's previous generation, and that the $2 and $10 introductory rates become the reference price other vendors answer. Broad self-serve access arriving quickly, with reproduced long-run agent results, would argue the opposite.

## What to watch

- Whether independent evaluations on long-horizon agent suites reproduce the 77.9% DeepSWE figure, which would tell you whether the lead is real or a function of Google's own harness.
- How fast Fairwind gating is lifted for ordinary API customers, which would show whether Google is prioritizing safety positioning or faster share capture.
- Whether OpenAI or Anthropic answers the $10 output price rather than the capability claim.