Google's Gemini 4 Carbon Tops Three Agent Benchmarks at Half the Cost

Google's strongest AI model spent its first two weeks in the hands of a small group of cybersecurity defenders. On Tuesday it went on sale to everyone else. Gemini 4 Carbon is now available to all paying developers and enterprises through the Gemini API, and rolling out to Google AI Ultra subscribers in the Gemini app and AI Mode in Search, at the same introductory price as Gemini 4 Argon: $2 per million input tokens and $10 per million output tokens. Google says it ranks first on three independent agent benchmarks at less than half the cost per task of the next best outside model.

Built to Think Longer Than Argon

Carbon sits above Gemini 4 Argon, which Google launched recently and still recommends for most coding and office work. Carbon is the model for the hardest jobs: multi-step coding and debugging, open-ended scientific research, long chains of agent work. It deliberately spends more tokens than Argon to explore, reason and check its own results in extended verification loops.

“Carbon is built for tasks that test the limits of a model's intelligence.”— Koray Kavukcuoglu, Google

Google announced the model on October 28 and gave it first to vetted defenders through its Fairwind Program. Two weeks later it reached the Gemini API in Google AI Studio, Gemini for business and Android Studio. Google AI Ultra subscribers also get it in Google Antigravity, the company's agent coding environment.

The price is the surprise. Carbon launched at Argon's introductory rate of $2 per million input tokens and $10 per million output tokens, with cached input at $0.10 per million. The introductory period ends on November 30, and from December 1 the rates double to $4 and $20. So how does a model that thinks longer come out cheaper?

Less Than Half the Cost per Task

Price per token and price per task are different numbers. A model that writes more but reaches a working answer in fewer attempts can cost less for the whole job. Google says Carbon ranks first on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench at more than 2X lower cost per task than the next best external model.

Google did not publish the underlying scores. VentureBeat reported Zapier's: Carbon completed 53.88% of AutomationBench tasks at an average of $2.12 each, while Anthropic's Claude Fable 5.2, at maximum effort, completed 51.45% at $4.64. On a new benchmark for agents that read dense files full of charts and blueprints, built with Artificial Analysis, Carbon ranks second at more than six times lower average cost than the leader.

The cyber version, still reserved for vetted defenders, shows what the extra reasoning buys. Scanning two major software-as-a-service providers for Wiz's Scan for Good program, it found previously unknown vulnerabilities in under two hours, compressing weeks of research into an afternoon, and Wiz worked with both companies to patch them. Google's own security system, Project Fortify, hit 100% recall on scanning and 95% precision on triage in internal tests.

The public model carries tighter guardrails. In simulated attack tests, its safeguards blocked 99% of the attack paths tried by an adaptive red-teaming agent and held the per-attempt breach rate to 0.8%, against 3.1% for Argon. Some of the most striking results, though, came from Google's own labs.

A Swarm of Agents and 40 Times More Clues

Google is also previewing /swarm in Antigravity, which turns a single prompt into a coordinated team of agents a developer can watch in real time. Carbon takes the roles that need the deepest reasoning, such as architecture, coordination and auditing, while faster models work in parallel and adversarial verifiers test every change. In one demo, a Carbon team took seven public city council planning PDFs and built an interactive 906,000-triangle 3D model of a Google office over 10 overnight loops.

In science, Google researchers used Carbon to hunt for the evolutionary relatives of proteins that structure predictors like AlphaFold 3 struggle with, because databases hold few close matches. Searching petabase-scale sequence archives across seven hard protein families from the CASP17 blind test, it found far more of them.

More relatives means more evidence for predicting how a protein folds, without retraining the prediction model. Drug-toxicity company Axiom Bio says Carbon is now its leading model for translational science, and Google is adding Carbon to its free program for scientists, which offers 10,000 researchers six months in Antigravity's highest usage tier.

Two weeks ago Carbon was a tool for a handful of security teams. Now anyone with a paid key can hand it their hardest problem.