Google on 10 November opened Gemini 4 Carbon — the model it calls its strongest — to all paying developers and enterprises through the Gemini API and to Google AI Ultra subscribers in the Gemini app, AI Mode in Search and the Antigravity coding platform. Two weeks earlier, Carbon had been limited to cybersecurity defenders in Google's Fairwind Program.
Carbon sits above Gemini 4 Argon in Google's frontier family. Google says Argon remains the better default for most software engineering and knowledge work, while Carbon is built for tasks that test a model's limits: multi-step coding and debugging, multi-package migrations and cross-subsystem root-cause investigations, and open-ended scientific research. It spends more tokens than Argon to "explore thoroughly, reason deeply, and run extended verification loops."
Alongside the release, Google is previewing /swarm in Antigravity, which turns a single prompt into a team of agents the developer can track in real time. Carbon takes the roles that demand the deepest reasoning — architecture, coordination and auditing — while faster, lighter models work in parallel in isolated workspaces, and adversarial verifier agents test every change. Google's examples include a 3D Eiffel Tower exhibit, a reconstruction of a Google office and a 3D pelican game, with one project run as ten overnight loops.
Pricing is the sharp edge. Carbon carries the same introductory rate as Argon — $2 per million input tokens and $10 per million output tokens — rising to $4 and $20 after 30 November. Google says Carbon ranks first on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench at more than twice lower cost per task than the next-best external model. Independent rankings are less flattering, and more revealing: on Artificial Analysis's Intelligence Index, Claude Fable 5.2 (max) scores 65, Carbon (high) 60, Claude Opus 5.5 (max) 59, GPT-6 Astra (max) 57 and GPT-6.1 Terra (max) 51 — but Artificial Analysis measured $7.76 per index task for Carbon against $26.54 for Fable 5.2.
The strongest cyber version stays gated: "Carbon for cyber" is available through the Fairwind Program, which defenders can apply to. Google says Carbon blocked 99% of attack trajectories from its adaptive red-teaming agent, cutting the per-attempt breach rate to 0.8% from 3.1% for Argon, and that external tester Trajectory Labs found no new universal jailbreaks across 132 hours and more than 13,000 requests. Google also repeats the result from the Signal work reported this week: AlphaProof Code agents used Carbon to generate millions of lines of Lean and prove key security properties of Signal's cryptographic protocol in weeks.
Technically, Carbon shares the same pre-trained sparse mixture-of-experts base as Argon, with extra post-training and reinforcement learning. It accepts text, images, audio and video with a context window of up to one million tokens, outputs text, and carries a March 2026 knowledge cutoff.
Sources
- blog.googleGemini 4 Carbon is now available — Google blog
- mlq.aiGoogle makes Gemini 4 Carbon widely available at $2 per million input tokens and previews agent swarms — MLQ.ai
- antigravity.googleGemini 4 Carbon and /swarm — Google Antigravity blog
- theverge.comGoogle releases Gemini 4 Carbon with agent 'swarm' feature — The Verge

