← Back to the record

Google DeepMind releases Gemini 4 Carbon — 59 points, first on AA Index v5

Date2026-10-28
CategoryFrontier Models
JudgementNeutral
Gauge weight2 / 3
SourceGoogle DeepMind

Summary

Gemini 4 Carbon, released by Google DeepMind on October 28, 2026, scored 59 points for first place on the Artificial Analysis Intelligence Index v5 (Opus 5.5 58, GPT-6 Astra 57, Argon 50). Company-reported cyber figures, evaluated without guardrails, were SEC-bench Pro 88.2%, ExploitGym 54.8% and CWE-bench 71.0%. Pricing was the same as Argon at $2/$10 (input/output per 1 million tokens), context was 1M, and it was initially released only to some verified cyber defenders through the Fairwind programme.

A/B Impact Analysis

This record is filed as neutral. The watch judgement treated it as a strong reconfirmation of the existing pattern and kept the median: the gain was +9 points over Argon, but the change at the frontier lead was only +1 point, the long-horizon claim was an anecdote with no success-rate figures, and on FrontierSWE 2 Carbon's 61.2% was below Astra's 65.5%. Honesty indicators such as 0.0% cheating success on Terminal-Bench 4.0, and the limited release to verified defenders, were also recorded.