Gemini 4 Carbon tops new Artificial Analysis Intelligence Index

Google released Gemini 4 Carbon, its next frontier model, just two weeks after the launch of Gemini 4 Argon. CEO Sundar Pichai stated on the earnings call that AI is increasing training and launch velocity, expecting continued acceleration over the next 12 months due to reinforcement learning. Carbon gained 8 points on Terminal Bench 4.0, 6 on Frontier Swee, 12 on Exploit Gym, 3 on CWE Bench, 7 on OS World, and 8 on Agent's Last Exam. It leads the new Artificial Analysis Intelligence Index V5 with a score of 59, a nine-point jump, ahead of Opus 5-5. On the AA Coding Index, Carbon is second, tied with Sonnet 5-5 and three points behind Opus 5-5. The cost per task is $7.76, which is 23% cheaper than Opus 5.5 and 20% more expensive than Astra. The model is currently restricted to cyber defenders and trusted partners under pre-release testing guidelines. Arena's Peter Gostead noted Carbon feels better than Argon, though render and game design tests were criticized for not matching Opus 5.5. DeepMind researcher Zachary Natto teased an unreleased model, while post-training researcher Xiao Zhang described the team as operating like a well-oiled machine. Artificial Analysis released Intelligence Index V5, which uses private datasets for the majority of benchmarks to prevent benchmark maxing. New benchmarks include Frontier Code 2.0 and Terminal Bench Science, while GDPVal AA, RSI code, and AA LCR were removed. The index equally weights knowledge work, coding, science, and general knowledge. Opus 5.5 leads with 58 points, followed by GPT-6 Astra at 57, GPT-51 Sol, Sonnet-5-5, and GPT-61 Terra. Gemini 4 Argon and Meta Wonderful tie for sixth. The gap between open-weight and US proprietary models widened from 3 to 12 points. Kimi K 3.1 Pro is the top open-weight model at 46 points. DeepSeek V41 Flash and GLM 5.3 lead released weights at 32 points, behind Claude Haiku 5.5 at 36. Costs vary, with GPT-61 Terra cheapest at $0.96 per task and Fable 5-1 most expensive at $13.04. A new Progress Index tracks model improvements since November 2022. Analysts noted the index better reflects real-world agentic performance, though debates continue over weighting and effort settings.