Gemini 4 Carbon is Google's new model for long-horizon agent work, built on Gemini 4 Argon but spending more thinking tokens on each task. It is now open to every paying Gemini API developer and in GitHub Copilot. Keep Argon as the default and send Carbon only the hard, multi-step tasks your evals show Argon failing.
What was released
Gemini 4 Carbon was announced by Google on 10 November 2026 as generally available, a fortnight after it first went to cybersecurity defenders through Google's Fairwind Program. Access now depends on the channel:
- Gemini API and Google AI Studio: all developers on the paid tier. The AI Studio link in the announcement uses the identifier
gemini-4-carbon. - Gemini for business and Gemini Enterprise: through the API, the model dropdown, Google Antigravity and CodeMender.
- Google AI Ultra subscribers: rolling out in Antigravity, the Gemini app and AI Mode in Search.
- GitHub Copilot: generally available on the Pro+, Max, Business and Enterprise plans, in the model picker across VS Code, Visual Studio, JetBrains IDEs, Xcode, Eclipse, Copilot CLI and the coding agent. The rollout is gradual.
The more capable cyber variant stays behind Fairwind Program verification. Google also previewed a multi-agent /swarm feature in Antigravity, which is not part of the API release.
What changed from Gemini 4 Argon
The Gemini 4 Carbon model card says Carbon is based on Argon and points to the Argon card for architecture, training data and acceptable use. The difference is how it works a problem: Carbon uses more thinking tokens than Argon to explore hypotheses, reason and run longer verification loops, with configurable thinking levels to trade quality against latency and cost.
Other facts from the model card:
- Text, image, audio and video input, text output, within a combined input and output context of up to 1M tokens.
- Knowledge cutoff of March 2026, less current in some domains.
- Google warns of occasional slowness or timeouts, and higher token use at higher thinking levels.
- Automated safety evaluations improved on Argon for text and multilingual content and dipped slightly for image-to-text.
On results, Google reports that Carbon ranked first on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench at a lower cost per task than the next external model. These are Google's own figures; run your own tasks before relying on them.
The price changes on 30 November
Carbon launches at the same introductory price as Argon for input and output tokens. When the introductory period ends on 30 November 2026, both rates double for the public Carbon model. Any cost test you run this month therefore understates the steady-state bill. Price your evals at the post-November rate, and remember that Carbon also emits more thinking tokens per task than Argon, so the gap per completed task is wider than the rate card shows. In Copilot, GitHub bills Carbon by usage at the provider's list price.
What it means for teams building AI agents
Keep Argon as the default. Google itself describes Argon as the right choice for most software engineering and knowledge work. Carbon is pitched at tasks that test the limits of a model: long debugging sessions, multi-agent engineering work and open-ended research. If Argon already passes your evals, Carbon adds cost and latency for no measured gain. Our Gemini 4 Argon migration notes still apply to the base model.
Route by task, not by team. The useful pattern is a router that sends a small class of hard tasks to Carbon and everything else to Argon. Define that class from evidence: take the tasks where Argon fails or needs several retries, run them on Carbon at each thinking level, and keep the lowest level that passes. Our notes on what AI agent evals catch cover how to build that test set.
Set timeouts and budgets before you switch. Longer verification loops mean longer turns, and the model card admits timeouts. Raise client timeouts for Carbon calls only, cap tokens per task, and alert on runs that blow through the cap. The approach in controlling AI agent running costs applies directly.
Decide the Copilot policy now. Under GitHub's default model enablement, new models switch on automatically unless an administrator has turned off the global default or this model. Business and Enterprise administrators who want Carbon spend reviewed first should set the model policy before the rollout reaches their organisation.
When not to switch. Stay on Argon for chat, retrieval answers, short tool calls and anything latency-sensitive. Wait if you rely on Vertex AI, since the announcement does not list it as a channel. And check that your knowledge needs fit a March 2026 cutoff, or keep retrieval in front of the model.
Frequently asked questions
What is the model ID for Gemini 4 Carbon?
The announcement's Google AI Studio link uses gemini-4-carbon, which points to that identifier for the Gemini API. Google's developer guide and what's new page in AI Studio give the full details, including thinking level settings. Confirm the exact ID in the AI Studio model list for your project before you hard-code it, and keep it in configuration so you can route between Carbon and Argon without a code change.
Is Gemini 4 Carbon more expensive than Gemini 4 Argon?
At launch it costs the same per token as Argon's introductory price. After 30 November 2026 the public Carbon rates double for both input and output tokens. Carbon also uses more thinking tokens than Argon on the same task, so cost per completed task is higher even while the rates match. Compare total tokens per finished task on your own workload, priced at the post-November rate.
Can I use Gemini 4 Carbon on the free tier of the Gemini API?
No. Google says Carbon is available to all paying developers through the Gemini API in Google AI Studio. Free tier projects will need billing enabled before they can call it. Google AI Ultra subscribers get it in consumer surfaces such as the Gemini app and Antigravity, and scientists can apply for time-limited free access through Antigravity's highest usage tier, subject to eligibility and capacity.
Will Gemini 4 Carbon turn on automatically in GitHub Copilot?
For organisations using GitHub's default model enablement, yes, unless an administrator has disabled the global default or this model specifically. It is available on Copilot Pro+, Max, Business and Enterprise, with a gradual rollout. Usage is billed at the provider's list price, so administrators who want to review cost first should set the model policy in Copilot settings before it appears for their developers.
