Gemini 4 Carbon Joins the Gemini API With a 1M-Token Window

Google has opened Gemini 4 Carbon to developers through the Gemini API and Google AI Studio, and the timing matters for teams that expect a model to finish what it starts. The introductory price only holds until the end of the month.

Built for longer verification loops

Gemini 4 Carbon is positioned as Google’s strongest reasoning model, built for sustained agentic workflows. The company lists complex software engineering, multimodal knowledge work, scientific reasoning, and cybersecurity defense as the areas it targets.

The coding pitch rests on verification rather than raw generation speed. Gemini 4 Carbon runs deeper build-and-verify loops to isolate root causes and test its own fixes against tests across large repositories, frontend interface design, and rendered 3D scenes. That matters in a repository large enough that one wrong assumption costs a day of review.

Google also promises tighter grounding on multimodal work. The model is built to draw fewer unsupported claims from dense PDFs, spreadsheets, slides, and workspace data, while holding closely to the output schemas a developer defines.

Cybersecurity defense sits in the same list of target workloads as software engineering. The pairing points security teams toward review and remediation work, not just writing new features.

One window for input, thinking and output

The context arrangement is the headline specification, because input, thinking, and output tokens now share a single 1,048,576-token window. A response can run past 64,000 tokens and up to the full window minus whatever the prompt consumed.

That design suits a model asked to explore before it answers. The reasoning Carbon does on a hard task draws from the same allowance as the answer it finally returns, instead of a separate budget.

Pricing, access and migration

Developers pay USD 2 per million input tokens and USD 10 per million output tokens until November 30, 2026. Once the introductory period ends, those rates double to USD 4 per million input tokens and USD 20 per million output tokens.

Access needs a paid Gemini API key or a Google AI Ultra subscription through Google AI Studio. Ultra subscribers also reach the model inside Google Antigravity, the company’s agent-first development environment. Google ships SDKs for Python, JavaScript, Go, Java, and C#, along with direct REST calls through cURL.

Migration support comes with the release: Google published a checklist and a skill that a coding agent can install to move an existing codebase onto the gemini-4-carbon model identifier.

Gemini 4 Argon arrived in the same channel in October under its own introductory rate. Gemini 4 Carbon sits above it in the lineup, and the two models answer different workloads rather than competing for one.

The introductory window is short, and teams that want to test Gemini 4 Carbon on a real workload have until the end of November to do it at the lower rate.