Gemini 4 Argon: API changes to make before you switch

Gemini 4 Argon is Google's new frontier model, generally available in the paid Gemini API as gemini-4-argon. Switching is more than a model string change: explicit context caching is gone, sampling settings are ignored and hard prompts can run for minutes. Budget for a price rise in December, and test thinking levels before moving production agents.

What was released

Gemini 4 Argon was announced by Google on 13 October 2026, in the Gemini API release notes and the developer announcement. The model ID is gemini-4-argon, released as a stable, generally available model. Google describes it as built for complex coding, agentic, maths and science work.

Developers can use it through the paid Gemini API in Google AI Studio and Android Studio. Google AI Ultra subscribers can also try it in AI Studio and Google Antigravity. On the same day, GitHub made it generally available in GitHub Copilot for Pro+, Max, Business and Enterprise users, billed at provider list pricing under usage-based billing.

What actually changed

The facts below come from Google's Gemini 4 Argon guide and model page.

• One shared context window. Input, thinking and output share a single window of 1,048,576 tokens.
• Long decode. With background execution and max_output_tokens above 65,536, one response can run past 64K tokens, generated in hops. Each hop adds a model_output step, so your code has to join the text across steps.
• Thinking takes more of the budget. Thinking levels are low, medium (the default) and high. max_output_tokens caps thinking and answer together, and Google advises raising it to at least 32,768 for apps that set a lower limit.
• Explicit caching is not supported. Calls that create or reference cachedContents must be removed. Implicit caching is on by default for prompts of at least 4,096 tokens.
• Sampling settings are ignored. temperature, top_p and top_k have no effect if sent.
• Slow on hard prompts. Google recommends background execution or client timeouts of at least 15 minutes.
• Tools. The model page lists function calling, structured outputs, code execution, search and Maps grounding, URL context, file search and Batch API as supported, and computer use in preview. The Live API, image generation and audio generation are not supported.
• Pricing. Standard input and output are introductory through 30 November 2026 at USD 2 and USD 10 per million tokens, rising to USD 4 and USD 20 from 1 December 2026. Thinking tokens are billed as output. It is not on the Free tier or the Google AI Pro plan.

What it means for teams building AI agents

In CodeDTX's view, the API changes matter more than any quality gain, because several of them fail quietly.

Ignored settings change behaviour without an error. If your extraction or classification pipeline relies on a low temperature for repeatable output, that control no longer exists on this model. Use structured outputs and validation to get consistency instead, and re-run your evals before trusting the new outputs.

Caching code needs rework. Teams that built cost control around explicit cached contents lose that lever and rely on implicit caching instead. Re-check what a long, shared system prompt actually costs per call; our guide to controlling what an AI agent costs to run covers which numbers to track.

Agent loops need longer timeouts and step handling. Gateways, serverless functions and queue workers often time out well before the model finishes a hard task. Move long tasks to background execution, poll for the result and make sure your parser joins every output step, or you will store half an answer.

Budget for the December price change now. Price tests run in October and November will understate the cost from 1 December. Model the post-introductory rate when deciding whether Argon is worth it for each route, and look at the Batch and Flex tiers for work that can wait.

Governance for coding assistants. In Copilot Business and Enterprise, new models are enabled automatically unless an administrator has turned off the default. Decide whether Argon fits your model policy before people start using it; our post on Copilot for JetBrains default models covers the related admin controls.

When not to switch yet

Hold off if your product depends on explicit caching, sampling controls or the Live API, or if your latency budget is measured in seconds. Routine, high-volume tasks are usually better served by a cheaper Flash model, with Argon kept for the hard cases. Also plan for change: our guide on what to do when a model is deprecated applies to the models you leave behind.

A sensible path is to route a small share of hard tasks to Argon, compare quality, cost and latency at all three thinking levels against your current model, and only then move whole routes.

Frequently asked questions

Is Gemini 4 Argon generally available or in preview?

Google lists gemini-4-argon as a stable, generally available model in its Gemini API release notes of 13 October 2026. It is available through the paid Gemini API in Google AI Studio and Android Studio. Some features around it are still marked preview, such as computer use, so check the model page for each capability your agents rely on before you depend on it in production.

Can I keep using explicit context caching with Gemini 4 Argon?

No. Google states that explicit caching is not supported on this model, so requests that create or reference cached contents need to be removed before you switch. Implicit caching is enabled by default for longer prompts and discounts repeated input automatically. Keep stable content such as system prompts and tool definitions at the start of each request so that implicit caching can match it.

Why are my temperature settings ignored on Gemini 4 Argon?

Google's migration guide says temperature, top_p and top_k are ignored if they are sent to this model. Requests still succeed, so nothing alerts you to the change. If your application relied on low temperature for repeatable answers, use structured outputs with a schema, validate every response in code, and re-run your evaluation set to confirm that outputs are still consistent enough for your use case.

How should we plan for the Gemini 4 Argon price change?

Google prices Gemini 4 Argon at an introductory rate through 30 November 2026, and the standard input and output prices double from 1 December 2026. Cost tests run before December will therefore understate the long-term spend. Model the later rate when deciding which routes to move, cap output tokens per request and consider the Batch or Flex tiers for work that does not need an immediate answer.