Google’s new Gemini 4 Carbon model debuted at No. 1 on the Text Arena with 1,528 points, No. 2 on the Agent Arena and No. 4 on the WebDev Code Arena with 1,768 points. In Agent Arena evaluations across more than 21,900 sessions, Carbon climbed from Gemini 4 Argon’s No. 7 spot with a 16.2% net improvement at a $1.80 median cost per task, 72% cheaper than the top-ranked Claude Opus 5.5 and ahead of GPT-6 Astra and Fable 5.1.

On the Artificial Analysis Coding Agent Index, Carbon ranked second with a score of 59 in Antigravity CLI, trailing Claude Opus 5.5 at 62 and beating Argon at 53, while placing second on both Terminal-Bench 4.0 at 64% and FrontierCode v1.1 at 51%. Standard pricing is set at $4 per 1M input tokens and $20 per 1M output tokens, matching Argon, with a 50% launch discount to $2 and $10. At standard rates, Carbon averages $7.76 and 123,000 output tokens per Intelligence Index task, 23% cheaper than Claude Opus 5.5 at $10.04 but 1.2 times the $6.25 cost of GPT-6 Astra.

Carbon also debuted at No. 2 on the Vals Index v2.2 behind Claude Sonnet 5.5 while taking No. 1 on both the Vals Finance Index and Legal Research Benchmark, using 2.3 times Argon’s reasoning tokens and 64% more output tokens. On AA-Omniscience v1.1, Carbon led all models with a score of 54 and 58% accuracy, while its hallucination rate rose to 11% from Argon’s 4%, remaining below the 50% rate recorded for both Claude Opus 5.5 and GPT-6 Astra. Announced two weeks after Argon with a 59 score atop the Artificial Analysis Intelligence Index v5.0 and a 54.8% score on the ExploitGym cybersecurity benchmark, Carbon is rolling out first to cyber defenders in Google’s Fairwind Program before broader availability.