Gemini 4 Carbon is now available

Two weeks after giving Carbon to a group of cybersecurity defenders through its Fairwind Program, Google is now releasing it to paying developers and enterprises via the API, and to Google AI Ultra subscribers in the Gemini app, AI Mode in Search and Google Antigravity. Google pitches it as its strongest model for long-horizon agent work, cyber, complex coding and scientific reasoning, with Argon remaining the default for most tasks.

The introductory price matches Argon at $2 per million input tokens and $10 per million output tokens. On November 30, it doubles to $4 and $20.

We covered Carbon's cyber debut as a story about staged access. Today's release is about something else: who gets to run the most expensive kind of thinking, and at what price.

Carbon is explicitly a token-hungry model. Google says it uses more tokens than Argon to explore, reason and run verification loops. Pair that with the new /swarm preview in Antigravity, which turns one prompt into a coordinated team of agents, and you get a product built for very large jobs. Google's showcase examples, like a multi-agent reconstruction of an office building from city planning PDFs across ten overnight loops, show both the ambition and the bill. Pricing at Argon's level for three weeks is an invitation to wire Carbon into pipelines before the price doubles. That's fair business. Teams should budget for the post-November rate, not the launch rate.

Google's cost claims are bold: number one 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, and second on a new multimodal agent benchmark it co-launched with Artificial Analysis at more than 6x lower cost. Those are vendor-reported, and one benchmark is new and Google-partnered. We'll wait for independent runs.

What we find genuinely encouraging are the safety and honesty numbers attached to a wider release. Google reports Carbon passed 92 percent of a 375-task agentic honesty benchmark, up from 70 percent for Argon, declining to invent missing experimental data 98 percent of the time. In red-team testing, it reports a per-attempt breach rate of 0.8 percent, down from 3.1 percent. For science users, refusing to fabricate data matters more than any leaderboard.

The downside is plain. A model with Google's strongest cyber skills is now in many more hands, and Google itself says prompt-injection robustness still needs work. The cyber-specialized variant stays behind Fairwind, which is right. But we'd like to see how misuse in the general release is tracked and reported over the coming months, not only measured before launch.

Google also reports early results from defenders and researchers: Wiz finding zero-day flaws in two major SaaS providers in under two hours, Google's own team proving two open-source bugs exploitable so maintainers could patch them, and a 40-fold average increase in homologs retrieved for hard protein families. Its free science program, which opened Argon to up to 10,000 researchers, now includes Carbon.