Gemini 4 Argon has a 1M-token output limit — but builders cannot switch to it yet

Google announced Gemini 4 Argon on September 30, 2026. It is a frontier model for long-running coding, knowledge-work, and defensive-cybersecurity workflows. But it is not a model most builders can select in Gemini API, Vertex AI, or an IDE today: Google says the initial rollout is to trusted cyber defenders through its Fairwind Program. (Source: Google’s announcement, 2026-09-30)

Key facts:

• Gemini 4 Argon has a 1 million-token maximum output limit. Google says this rises from 64K tokens in the prior generation. This is an output limit, not a published context-window specification. (Source: Google, 2026-09-30)
• The announced introductory price is $2 per 1M input tokens and $10 per 1M output tokens. Cached input is priced at a 95% discount. (Source: Google, 2026-09-30)
• Google reports 77.9% on DeepSWE v1.1. Treat this as a vendor-published benchmark result, not a substitute for an evaluation on your repository. (Source: Google, 2026-09-30)
• Google has not announced a public model ID or a general-availability date. It says broader release will start with paid API customers and Google AI Ultra subscribers after the early testing phase. (Source: Google, 2026-09-30)

What this means if you build with Gemini

Do not replace a production model string with gemini-4-argon. Google’s live Gemini API model documentation is the place to verify a selectable model and its limits. Until Argon appears there or in the service you use, the announcement is a roadmap and access signal—not a migration instruction.

Do not use the announced list price for a budget forecast yet. The $2/$10 rate is an introductory price attached to a future public release. Usage availability, quotas, regional support, caching behavior, and the model ID are not public API contract details at this stage. Keep your existing Gemini model explicitly pinned and record its completed-task cost before any future trial.

Prepare an evaluation harness now, rather than guessing from DeepSWE. Keep a small repeatable suite: representative repo tasks, expected test commands, token count, wall-clock time, and reviewer findings. When access arrives, run Argon beside your pinned baseline on the same harness. A 1M-token output budget may help long trajectories, but it can also make a failed agent run much more expensive if your stop conditions are weak.

For a production-ready baseline while you wait, use our Gemini 3.6 Flash guide and its Gemini 3.8 Flash update. The practical rule is simple: upgrade only once the model is documented as available in your actual endpoint, then measure completed work rather than assuming a benchmark transfers.

Sources

• Google — Gemini 4 Argon: our next era of frontier intelligence
• Google AI for Developers — Gemini API models
• Axios — Google unveils long-awaited Gemini 4
• Ars Technica — Google announces Gemini 4 Argon AI model, but you can’t use it yet