# Google Ships Gemini 4 Argon With a 1M-Token Output Ceiling and Gives Cyber Defenders First Access

Google reports Argon leads 12 of 18 disclosed benchmarks, including 77.9% on DeepSWE v1.1, while trailing GPT-6 Astra by 10.5 points on FrontierSWE v2.

- Published: 2026-10-01T06:20:51.871Z
- Canonical: https://polylog.news/ai/2026-10-01/google-ships-gemini-4-argon-with-a-1m-token-output-ceiling-a
- 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 announced [Gemini 4 Argon](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/) on September 30, positioning it for long-horizon software engineering, enterprise knowledge work, financial and legal analysis, and cyber defense. The headline architectural change is on the generation side rather than the input side: the model's output ceiling rises from 64,000 tokens to 1 million, which Google frames as room to produce hundreds of thousands of tokens inside a single trajectory instead of combining many separate, bounded calls.

The disclosed numbers are Google's own. Argon scores 77.9% on DeepSWE v1.1 against 74.2% for Claude Opus 5.5 and 74.1% for GPT-6 Astra, 51.3% on AutomationBench, 91.7% on LVBench for long-video understanding, and 84.2% on the longer GraphWalks evaluation against 71.8% for Astra and 66.8% for Opus 5.5. On Harvey's legal agent benchmark the gap is wide in relative terms but low in absolute terms, 19.6% for Argon against 5.4% and 3.8% for the two rivals, which says more about how unsolved that task family is than about the size of Argon's lead.

The losses matter as much as the wins. [The Decoder's tally](https://the-decoder.com/google-gemini-4-argon-closes-the-gap-with-openai-and-anthropic-but-doesnt-take-a-clear-lead/) finds Argon ahead outright on 12 of 18 benchmarks and tied on one, with Astra ahead on three and Opus 5.5 on two. Argon trails Astra by 10.5 points on FrontierSWE v2 (55.0% against 65.5%) and trails Opus 5.5 by nine points on Terminal-bench 4.0. No independent reproduction exists yet, and the test that matters most for the 1 million-token output claim, whether the model's output stays coherent across a few hundred thousand generated tokens, is not something a single benchmark table can settle.

Distribution is the unusual part. Rather than a general launch, Google is routing Argon first to vetted cybersecurity professionals through its Fairwind Program, as [VentureBeat reported](https://venturebeat.com/technology/google-unveils-gemini-4-argon-retaking-benchmark-lead-over-openai-and-anthropic-but-in-limited-release). Introductory pricing is $2 per million input tokens and $10 per million output tokens, with cached input discounted 95%, [per MarkTechPost's summary](https://www.marktechpost.com/2026/09/30/google-deepmind-unveils-gemini-4-argon-with-1m-output-tokens-for-coding-knowledge-work-and-cyber-defense/). That input and output pair matches what Anthropic charges for Claude Sonnet 5.5, so the per-token list price of frontier access is converging even as the capability claims diverge.

## What this means

The mechanism is output length as a product axis: a 1 million-token generation budget changes which agentic workloads fit inside one session, which shifts cost from orchestration code to tokens and favors vendors with cheap cached input. Google gains on the long-horizon coding and document-review workloads where Anthropic has held developer mindshare, and OpenAI keeps a defensible edge on the hardest software engineering benchmark. The gated release to cyber defenders is the newer signal: a lab that considers a model's offensive and defensive security capability material is now treating launch sequencing itself as a safety control, following the staged-access pattern Anthropic and OpenAI used for earlier cyber-relevant releases. The likelier outcome over the next few weeks is that general availability arrives with a usage policy specific to security tooling, and that independent testing confirms Argon's long-context advantage while narrowing its coding lead. Evidence against that read would be third-party runs showing the DeepSWE and GraphWalks margins collapse under non-Google scaffolds, which would mark the release as a packaging change rather than a capability advance.

## What to watch

- Whether Argon reaches general API availability and on what security-use terms, which will show how much of the gated launch was caution and how much was capacity.
- Independent reruns of DeepSWE v1.1 and Terminal-bench 4.0 outside Google's own harness, the cleanest test of whether the benchmark lead holds.
- Whether rivals answer the 1 million-token output ceiling with their own long-generation limits rather than with price cuts, which would show where labs think the next competitive edge lies.