OutYet reporting

Gemini 4 Argon’s API migration has a few sharp edges

Google’s documentation describes a million-token window and long outputs, but also calls for changes to caching, sampling parameters, and client timeouts.

OutYet Editorial Desk

Google’s Gemini 4 Argon documentation lays out more than a larger context window: it describes a unified limit of 1,048,576 tokens for input, thinking, and output, and says a single response can extend up to that limit minus the input. Longer outputs use background execution, with the service continuing generation in steps of up to 64K tokens. That gives developers a way to handle long tasks in one interaction, though the available output space is shared with the prompt and the model’s thinking.

The documentation identifies the API model as `gemini-4-argon`, with medium as the default thinking level and low, medium, and high as options. It recommends raising `max_output_tokens` to at least 32,768 because thinking consumes part of that limit. For outputs over 64K tokens, developers need background execution and a limit above 65,536; Google also advises setting client timeouts to at least 15 minutes for hard prompts. These are concrete integration requirements, not just benchmark claims.

There are compatibility changes to account for. Google says explicit context caching through `cachedContents` or `caches.create` is unsupported, while implicit caching applies to prompts of at least 4,096 tokens and discounts cached input on cache hits. The docs also say `temperature`, `top_p`, and `top_k` are ignored. Teams migrating from another Gemini model should check these settings and their cost assumptions rather than expecting existing request parameters and cache workflows to carry over unchanged.

Google lists introductory API prices through November 30, 2026: $2 per million input tokens and $10 per million output tokens, with cached input at $0.10 per million; the listed rates double starting December 1. The price and token ceiling make Argon worth evaluating for long-running tasks, but the docs warn that reasoning tokens count toward output limits and that difficult requests can take several minutes. Reuters reported the broader availability announcement and the introductory pricing, while the API guide supplies the operational details; neither establishes how Argon will perform on a reader’s own workload, so migration tests should measure latency, token use, and quality on representative tasks.

Related models

Sources