Google Releases Gemini 4 Argon API With 1M Token Output Limit

Google DeepMind has pushed Gemini 4 Argon into general availability, and the hook is hard to miss: a sparse mixture-of-experts model that can produce up to 1M tokens in a single response. That’s a huge jump from the older 64K output limit, even if the total input-plus-output window still tops out at 1M tokens. Feed it a giant prompt and the room left for generation shrinks. Simple arithmetic, brutal consequences.

Argon is aimed at work that tends to get messy and long: coding, finance, legal tasks, science and cybersecurity. It accepts text, images, audio and video, but it only outputs text. Google says this is a new model rather than a fine-tune of an earlier Gemini, and it was trained on TPUs with JAX and ML Pathways. The company hasn’t disclosed total or active parameter counts.

The pricing is aggressive at launch. Paid Gemini API users get intro rates of $2 per million input tokens and $10 per million output tokens, with cached input tokens discounted by 95%. Those prices rise after November 30, 2026. Access runs through Google’s own stack: the Gemini API, AI Studio, the Gemini platform, Gemini for business, Google Antigravity, the Gemini app and AI Mode in Search for Google AI Pro and Ultra subscribers. Cyber defenders also get a separate Fairwind Program path without cyber guardrails.

On paper, Google likes the numbers. Argon leads its own table on several knowledge-work and long-context benchmarks, including 77.9% on DeepSWE v1.1, 68.90% on Vals Index and 84.2% on GraphWalks across the 256K-to-1M range. But the picture is not spotless. It lands at 55.0% on FrontierSWE v2, 57.4% on Terminal-Bench 4.0 and 57.6% on Terminal-Bench Science 0.1, where rivals do better. Google also flags hallucinations and occasional slowness or timeouts.

The model card’s safety language is especially careful. Under Google’s Frontier Safety Framework, the company says it cannot rule out biological and cybersecurity critical capability levels, even as it says safeguards lower the risk. That’s the modern AI product launch in one sentence: bigger, cheaper, more capable, and still wrapped in a warning label.