Research|LLM: Gemini 4 Carbon Puts Google Back in the Lead as RSI Accelerates Model Development

Executive Summary

1. Google announced Gemini 4 Carbon on October 28, about two weeks after Argon became broadly available on October 13. On version 5 of the Artificial Analysis (AA) Intelligence Index, Carbon ranks first with a score of 59, ahead of Claude Opus 5.5 at 58 and GPT-6 Astra at 57; Argon scores 50. Pricing matches Argon: introductory rates of $2 per million input tokens and $10 per million output tokens, rising to standard rates of $4/$20. All token prices below follow this input/output convention. Carbon is initially being made available to a select group of cyber defenders, with access then expanding to more partners through the Fairwind Program. Google has not announced when developers and consumers will get access.
2. Google is back in the top tier. With models from the three leading labs separated by just two points on the AA index, we believe competition in model training has intensified further. Consistent with our earlier view, Alphabet CEO Sundar Pichai said on the company’s earnings call the same day that AI is helping Google speed up model training and releases. He expects recursive self-improvement (RSI) to drive further acceleration over the next 12 months. Our understanding is that Google used model soups in this cycle to accelerate model development.

Benchmarks: First Overall, Still Behind Opus 5.5 in Coding

Google benchmark comparison for Gemini 4 Carbon (selected)

• Carbon’s strengths are long-horizon tasks, long context and cybersecurity; it does not yet lead in software engineering or computer use. It leads on 15 of the 17 benchmarks in Google’s comparison. The two exceptions are FrontierSWE 2, where it trails Astra and Opus 5.5, and OSWorld-2.1, where it trails Opus 5.5.
• Third-party evaluations paint a similar picture. AA says Carbon does not yet lead on coding and knowledge-work evaluations. On its Coding Agent Index, Carbon running in the Antigravity CLI scores 59, below the score of 62 achieved by Opus 5.5 running in Claude Code. On LMArena, Carbon ranks first on the text leaderboard and second in Agent Arena, behind only Opus 5.5.
• At standard pricing, Carbon’s weighted average cost per task on the AA Intelligence Index evaluations is $7.76, about 23% lower than the cost for Opus 5.5 at its highest reasoning setting.

Carbon Scores Nine Points Above Argon as Training Competition Intensifies