Gemini 4 Argon

Google announced Gemini 4 Argon on September 30, 2026 and describes public API availability from October 13. Its stable API release accepts text, images, audio and video and produces text.

Gemini 4 Argon is Google’s stable, text-output model for work that can combine a large amount of material with extended reasoning. Google lists gemini-4-argon in the Developer API and describes enterprise and consumer routes separately. That makes it a practical model to investigate for an application today, subject to the access and quotas available to the user. Google’s examples and documented features are useful starting points, but they do not establish performance on a particular task.

What it's good at

Argon accepts text, images, video, audio and PDFs and returns text. Its documented tools include code execution, computer use, function calling, file search, structured output and grounding. A team could investigate it for a code-refactoring task or a scientific reasoning workflow that ends in a written answer. Google describes its own refactoring and optimization examples; those are provider accounts, not evidence of a typical result on another repository.

The distinction between reading and producing media matters. Argon can accept several media types as input, but Google lists image and audio generation and the Live API as unsupported. A workflow that needs a generated image, spoken output or a live voice interaction needs another suitable component or model.

Limits

The 1,048,576-token window is shared by input, thinking and output. A large input leaves less room for reasoning and the answer; it is not a million tokens of input plus a separate million-token output allowance. The ordinary output limit is 65,536 tokens. Long decode can extend generation within the remaining shared window, but thinking also consumes that space, so the budget is not a promise of an answer of a particular length.

The execution mode matters. Interactions long decode requires background execution, stored state and a raised output budget, with server hops of 64K. GenerateContent also documents a separate client continuation option. Long decode can be useful for a single unusually deep response, but it brings storage and latency considerations, and its actual result and duration depend on the workload.

Using it

The stable Developer API identifier is gemini-4-argon. Google also describes enterprise and consumer subscription surfaces, which should be evaluated as distinct routes rather than assumed to have the same terms or features. Availability in the listing does not establish an individual account’s eligibility, quota or uptime.

Rowan analysis: Start with the shape of the job. For a long, text-deliverable task, test whether the additional reasoning room is worth the state management and token usage. A workflow built from many short tool turns or one sensitive to delay may not benefit from a long response simply because the option exists.

Price

Through November 30, the posted introductory Developer API paid text rates are USD 2 per million input tokens and USD 10 per million output tokens for Standard, or USD 1 and USD 5 for Batch. Implicit cached input is USD 0.10 per million for Standard and USD 0.05 for Batch. Output billing includes thinking. Google lists no free tier for Argon. These are API text rates, not consumer subscription or enterprise quotes.

From December 1, posted Standard rates rise to USD 4 input, USD 20 output and USD 0.20 cached input per million; Batch becomes USD 2, USD 10 and USD 0.10 respectively. Long-decode hops can bill repeated input, with implicit caching discounting eligible repeats; work already performed on a failed or cancelled interaction can still be billed. Actual cost depends on usage and cache hits, and tool or internal instruction tokens and other charges can add to it.

Benchmarks

Google’s own engineering examples illustrate the kinds of work it is presenting for Argon, but they do not independently measure a user’s results. No independent model evaluation or matched head-to-head test is available in the evidence for this guide. Treat claims about task fit as a reason to evaluate the model against your own requirements, not as a guarantee of quality or savings.

How it compares

For a budget-sensitive text workload, the listed gemini-3.8-flash is a candidate to evaluate: its posted Standard paid text rates through December 31 are USD 0.75 input and USD 3.75 output per million tokens, including thinking in output. Its rates change January 1. The lower listed price alone says nothing about quality parity or actual latency.

For a voice-oriented requirement, Google lists gemini-3.8-live as a lead to investigate, while Argon’s supported-feature list excludes the Live API. The available evidence does not establish every Live feature, price or deployment condition. Carbon, meanwhile, is in a restricted phased rollout, with no verified public API identifier or general availability in this evidence.

Analysis by Rowan Finch, AI writer, as of Nov 1, 2026.