Google made Gemini 4 Carbon available to developers, businesses, scientists and consumers on November 10, 2026. The company says the model is designed for complex, multi-step coding and open-ended scientific research. It is also previewing /swarm in Antigravity, a feature that assembles a coordinated team of AI agents from a single prompt.
Longer verification cycles and agent teams
Compared with Gemini 4 Argon, Carbon uses more tokens to explore problems more thoroughly, reason through them and run longer verification cycles, according to Google. The company says Argon remains its choice for most software engineering and knowledge work.
In the /swarm preview, Carbon can take on roles such as architecture design, coordination or auditing. Lighter models can work simultaneously in isolated environments while verification agents test changes, and users can track the team’s progress in real time.
In a Google demonstration, a team of Carbon agents used /swarm to create an interactive 3D exhibition of the Eiffel Tower’s construction from historical photographs and engineering drawings. The team worked from a single prompt.
For web application development, Google says the model repeatedly builds and verifies its work, recovers from errors and checks both the code and the rendered user interface. The company also says Carbon analyzes large collections of documents and data and produces spreadsheets, presentations and reports following specified structures and formatting.
Results in agentic tests
Google says Carbon ranked first on CursorBench 4.0, Perplexity WANDR and Zapier AutomationBench. It placed second on Multimodal Knowledge Agent, a test Google developed with Artificial Analysis to assess multimodal performance in agentic tasks.
In pre-release evaluations, the model matched the best model tested on Hebbia E2E Financial Agentic, according to Google. It also ranked at the top among 23 model configurations on Balyasny Asset Management’s General Financial Intelligence benchmark.
In a 375-task agentic honesty test, Google says Carbon passed 92% of tasks, compared with 70% for Argon. Carbon declined to fill in missing experimental data in 98% of cases and flagged methodological flaws in 68%.
Cybersecurity and code verification
A specialized version of Carbon for cybersecurity is intended for defenders participating in the Fairwind Program. Google says it leads on SEC-bench Pro for finding vulnerabilities, ExploitGym for exploiting them and CWE-bench v1 for patching them.
In an internal benchmark for Project Fortify, Carbon achieved 100% recall (the share of relevant cases found) in scanning and 95% precision (the share of flagged findings that were valid) in triaging findings, according to Google. In tests of open-source repositories, the company reported a relative improvement in recall of more than 25% over Argon.
Google researchers also used Carbon alongside other models in AlphaProof Code to formally verify code underlying the Signal cryptographic protocol. Agents generated millions of lines of code in Lean and, according to the company, helped prove key security properties within weeks.
Protein research and access for scientists
In other Google research, Carbon increased the average number of related proteins known as homologs found across seven protein families from the CASP17 test set by 40-fold when used in a multi-agent setup. The company says broader benchmarking is still underway.
Google is extending its science program to Carbon. Selected current and future participants will receive six months of free access to Carbon and Argon at Antigravity’s highest usage tier. Applications remain open, with admission subject to eligibility, capacity and selection.
Where Gemini 4 Carbon is available
Paying developers and businesses can access the model through the API in Google AI Studio and Gemini for business. Its introductory price is the same as Argon’s: $2 per million input tokens and $10 per million output tokens.
Google AI Ultra subscribers can use Carbon in Antigravity, the Gemini app and AI Mode in Search. Business access also includes model selection in Gemini for business, Antigravity and CodeMender; cybersecurity defenders can apply to the Fairwind Program.
