On November 10, 2026 Google expanded Gemini 4 Carbon to paid API and Ultra users. This follows October’s restricted introduction rather than repeating that event.

The public API costs $2/$10 per million input/output tokens through November 30 and $4/$20 thereafter. Fairwind remains a separate defensive-security route.

Inputs and outputs have different boundaries

The model card specifies text, images, audio and video inputs, text output, and a combined input-output context of up to one million tokens. It also warns about hallucinations, slowness and timeouts. Higher thinking levels can consume more tokens. These are developer-described capabilities and limitations, rather than findings from a test conducted by this publication.

For readers, a large context provides potential room for more material in one task. It does not establish that every line will be interpreted correctly or that conflicting evidence will be reconciled. Our editorial assessment is that input capacity and the quality of the final conclusion need separate evaluation.

The cost of an acceptable result

Consider an illustrative change spanning several connected software packages. A useful result includes compatible interfaces, passing checks and a clear explanation. A lengthy answer that leaves one package incompatible fails the task, however persuasive its narrative. This is a suggested pilot criterion, not a measurement of Carbon performed by SIExpo.

A cost comparison should account for retries and review time. A cheaper token rate can be offset by generating more tokens; a longer verification process might be worthwhile if it prevents later repair. Neither conclusion follows automatically from the tariff without an equivalent task and equivalent conditions.

A practical pilot would retain the input, model version, thinking configuration and acceptance criterion. It could then compare cost, elapsed time and correctness together. Teams should separately establish what their actual account can access: a model name appearing in an application does not demonstrate that the same operating mode exists in its API.

That distinction matters when reporting success as well. A completed response, a change that passes checks and a change accepted by a maintainer are different observations. Recording them separately makes the spending meaningful and helps identify whether the bottleneck lies in generation, verification or human review. This is our proposed evaluation approach, not an additional product feature.

Broader availability allows more teams to make those comparisons. It does not turn company rankings into a universal reliability guarantee. SIExpo checked the announcement and documentation; it did not run benchmarks, security operations or paid requests for this article.

Sources: Google — Gemini 4 Carbon broader availability, Google DeepMind — Gemini 4 Carbon model card. The cover is a conceptual AI illustration, not a documentary photograph.

Editorial verification: 2026-11-11T17:08:13.509948+00:00.