Skip to content

GoogleFrontier

Gemini 4 Carbon

Google's strongest model. Open to cyber defenders from 28 October 2026 and to paying developers and Ultra subscribers from 10 November, at $2 / $10 per million tokens through 30 November. Top of LMArena's text leaderboard (1527) as of 9 November 2026.

Announced
Released
Access
Public

Benchmark scores

TestScoreBest on the trackerHow it was measured
AA Intelligence Index v5about ten tests in one score60 pointsClaude Fable 5.2 64.6 pointsIndependentv5.0.1; listed as Gemini 4 Carbon (High)Run 10 Nov 2026
FrontierSWE V2very long engineering jobs61%Claude Fable 5.2 70%Independentproximus harness, high effort, mean@5 over 34 tasks, 20-hour budget (best@5 66.7)
APEX-Agentsprofessional office tasks84%Top scoreIndependentpass@1, unknown effort; listed as "Gemini 4 Carbon"
ARC-AGI-3puzzle games, standard harnessTypical first-time human player: 100%65%Claude Fable 5.2 81%IndependentARC Prize verified, Standard harness, low reasoning (63.9% at medium; high not scored, 30 of 55 games finished). Provider Adapter harness: 99.1% at mediumRun 10 Nov 2026
ARC-AGI-2visual puzzles, model aloneAverage person: 60%Saturated94%GPT-6 Astra 95%IndependentARC Prize verified, semi-private set, medium reasoning (92.1% at high, 49.6% at low)Run 10 Nov 2026

Compare every model on the same tests →

Release story

Sources