Google opens Gemini 4 Argon API: up to 1M tokens of output per response

What happened

Google DeepMind has made Gemini 4 Argon generally available to developers, enterprises and paid Gemini subscribers. Google first announced Argon on September 30 for a small group of cyber defenders; the model is now reachable through the paid Gemini API.

The headline change: output length

Argon can write up to 1M tokens in a single response. Google says earlier Gemini models capped output at 64K. According to the model card, input and output share a combined 1M-token window, so a long prompt reduces the room left for output. Google also documents a new continuation protocol for long decodes.

Pricing and access

Introductory pricing is $2 per million input tokens and $10 per million output tokens. A full 1M-token response therefore costs $10 in output tokens. The model runs only on Google's cloud, via the Gemini API, AI Studio and the Gemini platform; no open weights are released.

What it can do

Argon is a sparse mixture-of-experts (MoE) model. It accepts text, images, audio and video as input and outputs text. Google says the knowledge cutoff is March 2026 and that the model was trained on TPUs using JAX. Its target domains are coding, finance, legal, science and cybersecurity.

What the benchmark table shows

Google compares Argon with GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5. Argon posts the top score on most rows, including 91.7% on LVBench for long video understanding and 77.9% on DeepSWE v1.1. It trails on some rows, though: 55.0% on FrontierSWE v2, last among the four models, and 57.4% on Terminal-Bench 4.0. The model card also lists hallucinations and occasional slowness or timeouts.