Google says it has a new best AI model, and once again, you can't actually use it. Gemini 4 Carbon arrived on Tuesday, just two weeks after Gemini 4 Argon, and Google is treating it less like a brand-new release and more like a sharpened version of what came before.
The company says Carbon makes its biggest gains in cybersecurity, coding and what it calls agentic tasks, meaning the model can carry out multi-step actions on its own rather than just answering a prompt. Google says Carbon shows a "significant leap" in finding, confirming and fixing software vulnerabilities, and that it's twice as likely as Argon to write code with no vulnerabilities at all. Chrome's team, for its part, claims Carbon is ten times better at generating what it calls "high-quality" security patches than rival models not built by Google.
On the WebDev Arena leaderboard, a site that ranks AI models on web development tasks, Google says Carbon's score jumped by 89 points, putting it at fourth place. Google's own charts have Carbon ahead of GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1 in nearly every test it ran. The margins, though, are thin. Every frontier model these days tends to land in the same few points of each other on these benchmarks, and Carbon's edge over the pack is no exception.
Google is leaning harder on security scores this time. In the Grey Swan evaluation, which tests how well a model resists prompt-injection attacks (attempts to trick an AI into ignoring its instructions), Carbon posted a 0 percent success rate for attackers. Argon scored 0.7 percent, Claude Opus 5.5 scored 1 percent, and GPT-6 Astra came in at 8.5 percent.
None of that matters much to anyone outside a research lab right now, because Carbon isn't out. Google is running the same staged rollout it used for Argon, citing a "deliberative approach" given Carbon's agentic coding and cybersecurity reach. The company says it's working inside the voluntary testing framework the US government has set up for frontier models, and plans to route Carbon through its Fairwinds cyber defense program before anyone outside Google gets near it.
When it actually ships, and what it costs
Google hasn't given a firm release date. Argon took roughly two weeks to go from announcement to limited availability, and Google isn't promising Carbon will move any faster. When it does land, AI Ultra subscribers and developers using the Gemini API will likely get it first.
On pricing, Google is repeating its Argon playbook: a discounted introductory rate of $2 per million input tokens and $10 per million output tokens through the API, doubling once the promotional window closes. Argon's discount runs through the end of November. Google won't say when Carbon's cheaper rate expires, which suggests the company hasn't settled on an end date yet, or isn't ready to commit to one publicly.
ADFiled by The AI Desk
Models, assistants and the companies and chips behind them, reported from what was released and what was claimed, with the difference kept clear.
More from this desk →
Be the first to comment
Join the argument. No password, just your email or a passkey.