A frontier model at an introductory price
Google made Gemini 4 Carbon generally available on Tuesday, describing it as “our strongest model with frontier performance in cybersecurity capabilities and long-horizon agentic workflows”. The announcement was written by Koray Kavukcuoglu, senior vice-president at Google DeepMind and the company’s chief AI architect.
Carbon sits above Gemini 4 Argon rather than replacing it. Google still points most software engineering and knowledge work at Argon, and positions Carbon for harder problems — complex multi-step debugging and open-ended scientific research — where it spends more tokens reasoning and verifying before it answers.
The pricing is the detail with a deadline attached. Carbon launches at $2 per million input tokens and $10 per million output tokens, identical to Argon. After 30 November 2026 the public-access model moves to $4 and $20, a doubling of both halves of the rate.

Paying developers and enterprises get it through Google AI Studio and Gemini for business. Google AI Ultra subscribers are receiving it in Google Antigravity, the Gemini app and AI Mode in Search. Cyber defenders had it two weeks before this launch through Google’s Fairwind Program, and scientists can apply on a rolling basis to an expanded free six-month access programme in Antigravity.
The scores are Google’s own
Google reports Carbon at number one on CursorBench 4.0, Perplexity’s WANDR and Zapier’s AutomationBench, at more than twice the cost-efficiency per task of the next best external model, and at number two on the Multimodal Knowledge Agent benchmark at more than six times lower average cost per task than the model above it. It says Carbon leads SEC-bench Pro, ExploitGym and CWE-bench v1 on finding, exploiting and patching vulnerabilities, and reports 100% recall on scanning and 95% precision on triage in Project Fortify, an internal benchmark.
None of those figures is an independent result. They are vendor-reported, and the Project Fortify numbers come from a benchmark Google both built and ran.
On the science side, Google says that across seven low-homolog protein families in the CASP17 blind test set, Carbon increased retrieved homologs forty-fold on average. Axiom Bio, a partner, is quoted calling it the leading model for translational science.
One prompt, a team of agents
Alongside the model, Google opened a preview of /swarm in Antigravity, which turns a single prompt into a coordinated team of agents the user can watch in real time. Carbon takes the high-reasoning roles — architecture, coordination and auditing — while faster models run tasks in parallel and adversarial verifiers test the changes.

Google also published safety figures for a model it is explicitly selling on cyber capability. It says Carbon blocked 99% of attack trajectories in simulated cyberattack tests, with a per-attempt breach rate of 0.8% against Argon’s 3.1%; that automated cyberattack query success in Gray Swan’s testing fell to 2.17% from Argon’s 9.86%; and that on a 375-task agentic honesty benchmark Carbon passed 92% of tasks against Argon’s 70%. Google notes that robustness still has room to improve.
What to watch
Two dates. The first is 30 November, when the introductory rate expires and Carbon becomes twice as expensive as the model Google says most users should be on anyway. The second is whenever an outside evaluator publishes numbers of its own: a model released to cyber defenders two weeks ahead of everyone else is a disclosure decision, and so far the only scores attached to it are the vendor’s.