Google releases Gemini 4 Carbon, its strongest model, to developers
Google made Gemini 4 Carbon generally available on November 10, 2026, two weeks after giving it to a small group of cybersecurity defenders. In its announcement, Google calls Carbon its strongest model, built for long-running agent work in software engineering, knowledge work, science and security. Developers can call it today through the paid tier of the Gemini API, at an introductory price that doubles after November 30.
What’s new
Carbon sits above Gemini 4 Argon in Google’s lineup. Google says Carbon spends more tokens exploring a problem and running verification loops, and still recommends Argon for most everyday software engineering and knowledge work. In Google’s words, “Carbon is our strongest model with frontier performance in cybersecurity capabilities.”
Where you can use it:
| Who | Where |
|---|---|
| Developers | Gemini API paid tier, in Google AI Studio and Android Studio |
| Enterprises | Gemini for business (API or model picker), Google Antigravity, CodeMender |
| Google AI Ultra subscribers | Gemini app, AI Mode in Search, Antigravity and AI Studio |
| Scientists | Free for six months on Antigravity’s highest usage tier, by application |
| Cyber defenders | “Carbon for cyber” through the Fairwind Program, by application |
The announcement does not mention Vertex AI or a context window size.
Pricing. Until November 30, 2026, Carbon costs $2 per million input tokens and $10 per million output tokens, the same as Argon. After that the price becomes $4 per million input tokens and $20 per million output tokens.
/swarm. The headline feature for developers is a preview of /swarm in Google Antigravity, Google’s agentic coding environment. It turns one prompt into a team of agents: Carbon handles the architecture, coordination and auditing, faster models do the parallel work, and adversarial verifier agents check the changes.
Benchmarks. Google says Carbon ranks first on CursorBench 4.0, Perplexity WANDR and Zapier AutomationBench at more than 2x lower cost per task than the next-best outside model, and second on the Multimodal Knowledge Agent benchmark run with Artificial Analysis. On an internal honesty test of 375 tasks it passed 92%, against 70% for Argon. In cyber work, Google says it leads SEC-bench Pro, ExploitGym and CWE-bench v1, which cover finding, exploiting and patching bugs.
Safety. Google reports that Carbon blocked 99% of attack trajectories from an adaptive red-teaming agent in simulated cyberattacks, with a per-attempt breach rate of 0.8% against Argon’s 3.1%. Outside testers at Trajectory Labs found no new universal jailbreaks in 132 hours and over 13,000 requests. Google also says Carbon had a 0% success rate for attacks on the Gray Swan indirect prompt injection benchmark, while noting that prompt injection robustness is still an area to improve. A model card and safety report are linked from the post.
What it means for developers
The model ID in AI Studio is gemini-4-carbon, and it needs a paid Gemini API key; there is no free tier for it in the announcement. Gemini 4 Carbon is also available outside Google’s own tools:
- Vercel AI Gateway lists it as
google/gemini-4-carbonwith a 1 million token context window, at 50% off through November 30, 2026, matching Google’s introductory price. - Devin added it to Devin Desktop and Devin CLI.
Vercel’s example with the AI SDK:
import { streamText } from 'ai';
const result = streamText({
model: 'google/gemini-4-carbon',
prompt: 'Write a TypeScript helper that safely parses JSON and returns a fallback.',
});The price change on December 1 matters if you plan to move agent workloads to Carbon: output tokens double to $20 per million, and Google itself says Carbon uses more tokens than Argon to reach its answers. Test on Argon first if your task is routine, and measure cost per task rather than per token.
Carbon is already behind some of Google’s research work. It powered the proof agents in DeepMind’s AlphaProof Code, which we covered in our post on the Signal protocol proofs.
How it compares
Every number above is Google’s own. Two outside data points so far:
- On the FrontierCode 2.0 coding benchmark, Devin reports Carbon at 40.0%, up from 26.3% for Gemini 4 Argon and slightly ahead of GPT-6.1 Sol at 39.0%, with a cost per task under two thirds of Argon’s at its top score.
- Anthropic’s launch table for Claude Fable 5.2 a day earlier listed Gemini 4 Carbon ahead of Fable 5.2 on AutomationBench (53.9% against 51.4%), but behind it on most other rows. See our post on Fable and Mythos 5.2.
Independent leaderboard results for the public model were not out at the time of writing.