Timeline

Google makes Gemini 4 Carbon generally available

Priced as Argon at $2 and $10 per million tokens; Google's new safety report says it cannot rule out the cyber and biology critical capability levels without mitigations.

  • Models & capabilities
  • Safety & alignment
  • Notable

Google made Gemini 4 Carbon available to developers, enterprises, scientists and consumers, two weeks after it released the model only to cyber defenders through its Fairwind Program. That restricted first stage is why general release is recorded here, as it was for Argon. Gemini 4 Carbon is in Google AI Studio and Gemini for business at the same introductory price as Gemini 4 Argon, $2 per million input tokens and $10 per million output tokens, and is rolling out to Google AI Ultra subscribers in Antigravity, the Gemini app and AI Mode in Search.

Google said Carbon ranks first on CursorBench 4.0, Perplexity’s WANDR and Zapier’s AutomationBench at what it called more than 2X lower cost per task than the next best external model, and second on a Multimodal Knowledge Agent benchmark it built with Artificial Analysis. These are Google’s reading of other organisations’ leaderboards. It also described AlphaProof Code, which used Carbon among other models, and a /swarm preview in Antigravity.

The same day Google published a safety and security report for Carbon; the October launch post had pointed to Argon’s. Under its Frontier Safety Framework, Google said Carbon’s general capability rose by less than its threshold for a “material capability change” over Argon, so it ran no full critical-capability assessment and carried Argon’s determinations forward. That means it cannot rule out that the model, without outside mitigations, reaches the Cyber Uplift Level 1 or CBRN Uplift Level 1 (biological) critical capability levels, or the stealth and situational awareness tracked level. It rules out the chemical, radiological and nuclear level and the harmful-manipulation and ML R&D levels. Google said safeguards bring the residual risk to an acceptable level, and that because ExploitGym scores rose (54.8 per cent, 12.8 points above Argon) it hardened its cyber protections.

The report also set off a dispute. It says RL rewards are meant to judge outputs rather than thought content, but that learned components assigning token-level credit see the model’s thoughts. Redwood Research wrote that a similar passage in the Gemini 4 safety report suggested Google trains on chain of thought, which it called concerning, while Google’s own blog says it does not penalise the model for what it writes there. Redwood said Google staff had told it privately they were not training on chain of thought and asked Google to clarify. Google had not replied in the sources reviewed.

In the commentary

What people were saying around this time — external links, from the record's commentary rail.