Gemini 4 Carbon: API Access, Announced Pricing, and What Benchmarks Show

If you are evaluating Gemini 4 Carbon, first check whether your project can actually use it. A model announcement, an API catalog entry, and account-specific access require separate checks. This article reviews Carbon's availability and announced prices, then turns benchmark results into a practical evaluation plan for development work.

An AI-generated conceptual illustration of evaluating a new model in a controlled environment. It is not a product screenshot or a depiction of the model's internals.

Documentation checked: November 2, 2026. This article reviews public sources; it does not report hands-on Carbon API calls or model performance tests.

1. Separate the announcement from public API availability

Google announced Carbon on October 28, describing initial access for selected cyber defenders and Fairwind partners. Broader developer, business, and consumer availability was presented as a subsequent step. Google's announcement

On November 2, the Gemini API model catalog listed gemini-4-argon under Gemini 4; I found no Carbon entry. That observation does not establish the access status of every private partner.

For an adoption ticket, record the available platform, exact model ID, and access approval status alongside the announcement link. Guessing a model identifier and changing configuration first makes access failures harder to distinguish from code defects.

2. Separate announced prices from actual billing terms

Carbon's announcement gives the following rates, in USD per million tokens.

These figures come from the announcement and its pricing footnote. Do not automatically apply Argon's December 1 price change in the current API pricing page to Carbon. Carbon's announcement does not specify when its introductory period ends.

For example, one million input tokens plus 100,000 output tokens, without caching, would cost $3 at the introductory rates and $6 afterward. This is arithmetic using the announced rates, not a billing quote.

In a workload cost sheet, track retries, tool calls, rework after failed tasks, and human review time separately from token rates. Comparing the cost of a completed result gives a more useful basis for an operating budget.

3. Do not read a leading index score as your workload's success rate

The Artificial Analysis Intelligence Index v5.0 combines multiple evaluations. At the time checked, Carbon High led with 59 points. That does not mean a 59% task success rate in your service.

Selected rows from Google DeepMind's published table show why the type of work matters.

Source: Google DeepMind's performance table, accessed November 2. Rankings and values may subsequently change. Carbon leads the terminal-work row above, while Opus has the highest OSWorld score. Focus on evaluations close to the work you need done.

The official evaluation methodology describes the highest thinking settings unless otherwise specified and supplier-reported results for some competing models. OSWorld uses an 82-task offline subset; the Gemini scores report the maximum over three runs, with a single attempt per run.

This table should therefore not be presented as a fresh evaluation of every model in one shared internal environment. Benchmark versions, tools, time limits, and retry policies can all change practical results.

4. Build an evaluation set while waiting for access

The following is this article's adoption proposal, not Google's official usage procedure. Establish a baseline with your current model, then compare Carbon under the same conditions once access is available.

For example, “make this failing test pass for a specific input while preserving the public function's behavior” is easier to compare than “fix the bug.” Review whether the model changed unrelated files or weakened tests to obtain a pass.

Keep the model's completion report separate from the actual verification logs. A report containing a test command is different evidence from a log showing that command ran and passed. Setting criteria beforehand helps prevent a polished demonstration from driving a premature production switch.

Start by confirming an official access route and recording a baseline for the current model. Once access opens, use the same evaluation set to compare quality, elapsed time, and cost per completed task. That produces evidence for an actual migration decision.

한국어 버전: Gemini 4 Carbon 공개: API 접근·가격·벤치마크에서 확인할 것