Kimi K3.1 beats GPT-6 Astra in software engineering marathon

Moonshot AI's open-weight Kimi K3.1, a 3-trillion-parameter model, outperformed GPT-6 Astra and Opus 5.5 in the Software Engineering Marathon and scored 92.7 on the Bench CAD test. Open weights arrive November 4, 2026, with task costs of $2 input and up to $10 output, roughly 80% cheaper than GPT-6 Astra.

• 3-trillion-parameter model beat GPT-6 Astra and Opus 5.5 in coding
• Scored 92.7 on Bench CAD and strong results in Frontier Finance
• Pricing: $2 per task input, up to $10 output
• Open weights release on November 4, 2026