Google makes Gemini 4 Argon generally available with 1M-token output

Google first announced Argon on Sept. 30 for a small group of cyber defenders. The model targets coding, finance, legal, science and cybersecurity and is designed for long, multi-step work.

According to the model card, Argon is a new model rather than a fine-tune of an earlier Gemini. It is a large-scale sparse MoE transformer with native multimodal support. It accepts text, images, audio and video and outputs text. Google trained it on TPUs using JAX and ML Pathways. The knowledge cutoff is March 2026. Google has not disclosed total or active parameter counts. The model runs only on Google's cloud, through the Gemini API, AI Studio and the Gemini platform, and no open weights were released.

Google raised the output limit to 1 million tokens, up from 64,000 on earlier Gemini models. The model card describes a combined input-plus-output window of up to 1 million tokens, so a long prompt reduces the space left for output. Google documents a new continuation protocol for long decodes in developer and enterprise documentation. At introductory pricing, a full 1 million-token response costs $10 in output tokens.

Google compares Argon with GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5. It posts the top score on most rows of Google's table. Argon scored 77.9% on DeepSWE v1.1, ahead of Opus 5.5 at 74.2% and GPT-6 Astra at 74.1%. It scored 68.90% on the Vals Index, ranked first by Vals AI ahead of Claude Sonnet 5.5 at 67.04%. On AutomationBench it scored 51.3%, against 42.5% for Opus 5.5. On Harvey's Legal Agent Benchmark it scored 19.6%, nearly three times the next model but still a low absolute score. On GraphWalks from 256K to 1M tokens it scored 84.2%, against 71.8% for GPT-6 Astra. On CWE-bench v1 it scored 68.0%, tied with GPT-6 Astra. Google also reports 87% on the ERQA embodied reasoning benchmark. Google's reported results place Argon at the top of its table on knowledge work, long context and video, with a best score of 91.7% on LVBench for long video understanding, while its agentic coding results are mixed.

Argon trails on five rows of Google's own table. It scored 55.0% on FrontierSWE v2, versus 65.5% for GPT-6 Astra, and was last among four models in Google's table. It scored 57.4% on Terminal-Bench 4.0, versus 66.4% for Opus 5.5; 57.6% on Terminal-Bench Science 0.1, versus 68.1% for GPT-6 Astra; 45.3% on PostTrainBench, versus 49.3% for Opus 5.5; and 69.2% on the offline subset of OSWorld-2.0, versus 72.6% for GPT-6 Astra.

MarkTechPost reported that readers should review the evaluation methodology before drawing conclusions. Most rival scores are self-reported or pulled from public leaderboards. Argon's DeepSWE score is self-computed with a mini-swe-agent harness, while rival DeepSWE numbers come from system cards and a leaderboard.

The model card lists hallucinations and occasional slowness or timeouts. Under Google's Frontier Safety Framework, Google says it cannot rule out the biological and cybersecurity critical capability levels. It reports deploying safeguards it assesses as lowering that risk.