Arena: on long-horizon Agent Arena tasks, Claude Opus 5.5 works longest and Kimi K3.1 Pro shortest

Arena, the leaderboard publisher, says long-horizon tasks in its Agent Arena, which run hundreds of turns (porting a codebase to Rust, auditing an investment strategy document, writing a math textbook), widen the differences between frontier models.

Comparing the top model of five labs (Claude Opus 5.5, Gemini 4 Carbon, GPT-6 Astra, Muse Wonder Full and Kimi K3.1 Pro), Arena reports a 55-minute gap in time spent on long-horizon tasks. Claude Opus 5.5, #1 in Agent Arena, takes the longest at every task length, up to 1 hour 51 minutes; Kimi K3.1 Pro takes the least, 56 minutes on long-horizon tasks.

Arena also says Google DeepMind and Xiaomi gain most as tasks get more complex. On cheaper tasks OpenAI and Anthropic hold 10 of the 11 frontier spots; on the most expensive 5% of sessions they hold 7, and Gemini 4 Argon and Xiaomi's MiMo-V2.6 Flash and Pro take the others. Arena adds that more than 1,000 sessions in the past two weeks cost $100 or more. The figures are Arena's own.