Google’s Gemini 4 Carbon has taken the top spot in Artificial Analysis’s newly revised Intelligence Index, with a score of 59, according to Kingy AI. The result is a notable early showing, but the comparisons also show Carbon trailing rivals on some tasks, while broad access remains unavailable.
Google DeepMind Watch analysis
What happened
Kingy AI’s Carbon guide says Artificial Analysis gave Carbon a score of 59 on its v5.0 index, one point ahead of Claude Opus 5.5 Max and two ahead of GPT-6 Astra Max. The guide notes that the index’s evaluation mix changed from v4.3, so scores across those versions are not directly comparable.
The same account says Carbon scored below rivals on some individual tests: Google’s chart puts Opus and Astra ahead on FrontierSWE2, and Opus ahead on the cited OSWorld computer-use result. Kingy AI also reports that access is initially limited to selected cyber defenders and Fairwind partners, with no date announced for wider availability.
Why it matters
A single leaderboard position can make a model look like the obvious winner. The task-level results in this account are a useful reminder that coding, computer use and long-context retrieval measure different things, and that an aggregate rank may not predict the best fit for a particular job.
For developers and buyers, the practical next step is not to shop by rank alone. Carbon’s broader availability has not been dated, and a headline score cannot replace testing the tasks, tools and costs that matter in a real deployment.
Our read
Carbon’s reported result is worth attention, but it is a lead on one revised index, not a universal verdict. Kingy AI says it has not tested the model itself; its guide draws on evaluator results and Google’s launch chart. That distinction matters, particularly when the reported gaps are small and other tests produce different leaders. The leaderboard is a starting point, not a procurement decision wearing a medal.
What to watch
- When Google opens Carbon to developers and businesses, and what the final service terms include.
- How Carbon performs on independently run tests that match real coding and computer-use workflows.
- Whether its narrow lead on the revised index holds as more results arrive.
Discussion spark: When model rankings disagree across tasks, should buyers start with the highest overall score or the tests closest to their own work?
Sources and evidence
- Gemini 4 Carbon: Specs, Benchmarks, Pricing and Availability – Kingy AI (28 October 2026, 19:46 UTC)
not affiliated with, endorsed by, or operated by Google or Google DeepMind