Moonshot releases Kimi K3.1 Pro model via API

Summary

Moonshot released the Kimi K3.1 Pro model via API on October 23, 2026, with a commitment to provide full open weights by November 4. The model scored 55 on Artificial Analysis, trailing only Claude Opus 5.5 (58). It is offered at a cost of $2 per million input tokens and $10 per million output tokens, which is half the price of Opus 5.5.

Why it matters

It is the first time a model slated for open-weight release has achieved near-frontier benchmark scores. This could erode the price advantage held by US-based closed-model labs. Offering the same price as GPT-6.1 Sol while being 50% cheaper than Claude Opus 5.5 places concrete commercial pressure on frontier models. If the open-weight commitment is fulfilled, developers will be able to host a model previously accessible only through proprietary providers on their own infrastructure.

Highlights

• The first near-frontier model (AA score 55) planned for full open-weight release (November 4, 2026)
• $2 per million input / $10 per million output tokens: Same as GPT-6.1 Sol, half the price of Claude Opus 5.5
• Highspeed layer is 2-3x faster but 2x more expensive (excluding input cache hits)

Critical take

A significant portion of the highlighted benchmark scores belongs to Moonshot. The difference of 55 versus 58 on Artificial Analysis is narrow enough that the ranking could change with a different evaluation set. Additionally, the model is not yet open-weight. The 2x price increase in the Highspeed layer indicates that the claim of being 'half the price of Opus' applies only to the base speed, not to the faster variant required by most production workloads.