Google opens Gemini 4 Carbon to developers with API prices set to double in December

Developers can now put Gemini 4 Carbon to work without qualifying for Google’s restricted cybersecurity program. Google opened general availability on November 10, 2026, bringing the model to its paid API and AI Studio. The opening comes with temporary pricing: input and output rates double after November 30, leaving a short window to evaluate it at introductory rates.

Access moves beyond cyber defenders

Carbon first appeared on October 28, when selected defenders received limited access through Google DeepMind’s Fairwind program. This release opens the general-use model more widely; the specialized cybersecurity version remains on a separate Fairwind access path.

The coordinated rollout also includes Gemini for business, Android Studio and Antigravity. Google says Carbon is reaching AI Ultra subscribers in the Gemini app and AI Mode in Search. Paid API access is available in supported countries and territories, although Google has not published a country-by-country rollout schedule.

Access does not mean ownership of the model. Carbon is a proprietary hosted service, with no downloadable weights or self-hosting license announced. VentureBeat also notes that the rollout leaves Vertex AI availability, regions and support commitments unspecified—details that matter for buyers tied to a particular cloud contract.

More reasoning, a shared token budget

Google positions Carbon for extended coding, scientific reasoning, cybersecurity defense and work combining different kinds of information. Its developer documentation describes a deliberate tradeoff: compared with Gemini 4 Argon, Carbon spends more thinking tokens exploring possibilities and checking its work. That increases both token use and response time.

Developers can choose three thinking levels at the same per-token price. Google describes the choices this way:

• Low uses the fewest tokens and responds fastest, but quality can fall on difficult tasks.
• Medium is the default and Google’s recommended starting point for most tasks, including complex coding and agent workflows.
• High uses the most tokens and takes longest to begin answering, targeting the hardest reasoning, science and coding work.

The model’s 1,048,576-token context window is also a shared budget, not separate million-token allowances. Input, internal thinking and the final answer all occupy that space. Google advertises responses approaching one million tokens, but the input and reasoning consume room that would otherwise be available for the answer.

Long responses require extra setup. Standard synchronous requests stop at 65,536 tokens. The configured maximum counts thinking and output together, so reasoning can leave less room for the final response.

Carbon’s API rates after the promotion

Uncached input

$2→$4

USD per million tokens

The introductory input rate applies through November 30, 2026; the higher rate follows.

Output

$10→$20

USD per million tokens

Output pricing doubles on the same schedule, from the introductory rate to the standard rate.

Cheap tokens do not guarantee cheap tasks

At introductory rates, Carbon’s input and output prices are one-quarter of Claude Fable 5.2’s standard rates. After the promotion, they will be half. But a model that spends more tokens reaching an answer can erase some of that apparent advantage.

Artificial Analysis’s estimates illustrate the difference. At published standard rates, Carbon on high reasoning averages $7.76 per task on its intelligence-index suite, versus $26.54 for Fable at maximum effort. At medium effort, however, Fable averages $3.80 with an index score of 56, while Carbon averages $6.59 and scores 58.

Those effort settings are vendor-specific, not interchangeable controls. Google’s own guidance is to test Carbon at all three levels and choose the lowest that meets the required quality. It also recommends Argon for most software engineering and knowledge work, reserving Carbon for tasks that need deeper exploration and verification rather than making it the automatic replacement.