Gemini 4 Carbon

Runs MCP tools

Google’s strongest model, generally available from 10 November 2026. It is built for long agent runs: multi-step coding, debugging and research, where it spends extra tokens checking its own work. Google still recommends Argon for most everyday tasks.

Vendor
Google DeepMind
Released
10 November 2026
Context
1.05M tokens
Input
Text, image, video, audio, file

Paste a server URL, pick a model, and watch it call your tools in a real conversation. No install.

No server of your own? Leave it blank and use one of the public mock servers.

In your browser

MCP Agent Studio

Server has a public URL? Chat with Gemini 4 Carbon on it, watch every tool call, and compare models side by side.

Open MCP Agent Studio →

On your computer

MCP Playground Desktop

Server runs on localhost or over stdio? The browser can’t reach it. The desktop app can, with Gemini 4 Carbon in Agent Studio. Download free

MCP Evals

Can Gemini 4 Carbon actually complete tasks with your server?

One chat shows the tool calls work. An eval writes real tasks from your tool schemas, has Gemini 4 Carbon drive each one, and reports where it got the wrong answer even though every call succeeded. Pick Gemini 4 Carbon as the driver.

What Gemini 4 Carbon is

Gemini 4 Carbon sits above Gemini 4 Argon in Google’s line-up. Google first gave it to cyber defenders through its Fairwind Program, then opened it in the Gemini API, Google AI Studio and Gemini for business. It has a 1M-token window, takes text, image, video, audio and file input, and offers low, medium and high reasoning effort on OpenRouter. Google’s own guidance is that Argon stays the better pick for most software and knowledge work. Carbon is for harder, open-ended jobs where extra exploration and verification are worth the tokens.

What Google DeepMind says

  • Ranked first on CursorBench 4.0, Perplexity’s WANDR and Zapier’s AutomationBench, at more than 2× lower cost per task than the next external model.
  • Gray Swan indirect prompt injection: 0% attack success. Google adds that injection robustness remains an area for continued improvement.
  • Agentic honesty, 375 tasks: passed 92%, against 70% for Argon.
  • Adaptive red-teaming on ExploitBench and CyScenarioBench: per-attempt breach rate 0.8%, down from 3.1% on Argon.

These are the vendor’s own claims, not measurements of ours. Run the model against your own server to find out whether they hold for your tools.

How it reaches MCP

  • MCP Playground’s Agent Studio: paste your server URL above and run Gemini 4 Carbon against it in the browser, with no Google key
  • The Gemini API in Google AI Studio, whose Gen AI SDKs can pass an MCP session to the model as tools
  • Google Antigravity, where Google AI Ultra subscribers get Carbon and a /swarm preview that coordinates several agents

What to watch for

  • Google says Carbon uses more tokens than Argon because it explores and runs verification loops. On an MCP server that can mean extra read calls before it acts. Budget for longer runs, and compare call counts with Argon on the same task.
  • Google’s launch pricing is introductory and ends on 30 November 2026. If you plan a long eval run, check the current price first.
  • Google calls Argon the better choice for most software and knowledge work. If Argon already passes your evals, Carbon may cost more without changing the result.

Run the same multi-step task on Carbon and Argon against the error-simulation mock, with a step that hits simulate_partial_success. Google reports a large jump in agentic honesty. Check which model names the step that failed, and how many extra calls Carbon makes before it answers. All the mock servers →

Available here

ModelModel ID
Gemini 4 Carbongoogle/gemini-4-carbon

Go deeper

Frequently asked questions

Does Gemini 4 Carbon support tool calling?
Yes. OpenRouter lists tools, tool_choice and structured_outputs, and Google built it for long agent workflows. MCP servers work through any client that turns MCP tools into function calls, including the Gemini SDKs.
Gemini 4 Carbon or Gemini 4 Argon?
Google recommends Argon for most software and knowledge work, and Carbon for harder multi-step coding, debugging and research. Carbon spends more tokens per task. Run both against your server and compare the results and the number of calls.
Is Gemini 4 Carbon safe to point at tools that return untrusted content?
Google reports 0% attack success on Gray Swan’s indirect prompt-injection benchmark, and also says injection robustness still needs work. Treat it as a strong vendor result, not a guarantee, and keep permission checks on tools that write or spend.
What reasoning levels does Gemini 4 Carbon have?
OpenRouter exposes low, medium and high reasoning effort. Higher effort means more thinking tokens before each tool call, so expect slower turns on long runs.

Sources