Google's Gemini 4 Argon opens first to trusted cyber defenders

Google DeepMind has launched Gemini 4 Argon, its new frontier model, but most people can’t use it yet. On September 30, 2026, the model rolled out only to a set of trusted cyber defenders in the Fairwind Program. Google says it will open access to developers, enterprises and consumers “as soon as possible,” after it tightens guardrails based on early testers’ feedback.

The change that matters most for builders is the output limit. Google says Argon now produces up to 1M output tokens, up from 64K. Artificial Analysis reached that number through Long Decode Continuation, a new API feature that pauses a long response and resumes it across several calls. Vals lists a maximum of 262K output tokens, so the 1M figure depends on how you count it.

Pricing starts at $2 per million input tokens and $10 per million output tokens, and cached input costs 95% less. Artificial Analysis says standard pricing is $4 and $20, and the introductory discount has no end date yet.

Google reports strong benchmark results. Argon scores 77.9% on DeepSWE v1.1 and 91.7% on LVBench, and Google says it leads the Vals Index. Artificial Analysis puts it at 53 on its Intelligence Index, level with GPT-6 Astra. A task costs $1.99 at the discounted price, against $3.26 for Astra, but Argon writes more: about 62K output tokens per task against 27K. The Google numbers are Google’s own. The rest come from outside evaluators with their own setups, so check them against your own work.

On security, Google will release Argon without cyber guardrails to trusted defenders and its internal teams. Wiz uses it through its Scan for Good program and says the model found a critical flaw that exposed personal data in healthcare software. Some observers questioned the published numbers. @BlackHC pointed out that Argon’s 19.6% on Harvey’s legal benchmark trails the 25.42% listed for Muse Spark 1.2, and @teortaxesTex raised concerns about benchmark tuning.