Kimi K3.1 Pro, Gemini 4 Carbon Top Terminal-Bench

This is a meaningful leaderboard upset, but not proof that either model universally out-codes Claude; harness and evaluation methodology still matter.

• –Terminal-Bench 4.0 tests 66 difficult terminal tasks and requires every task-specific test to pass.
• –Kimi reports 65.7% using the Claude Code harness, while Artificial Analysis reports 67.2% with mini-swe-agent, underscoring the importance of apples-to-apples comparisons.
• –Kimi K3.1 Pro offers a 1-million-token context window and API pricing of $2 per million input tokens and $10 per million output tokens. [Kimi](https://www.kimi.ai/blog/kimi-k3-1-pro)
• –Gemini 4 Carbon matches that introductory API pricing but is initially limited to selected cyber defenders rather than broadly available to developers. [Google DeepMind](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/)
• –The bigger story is competition compressing rapidly: agentic coding leaders are changing within days, not quarters.