What changed
On October 23, 2026, Moonshot AI launched Kimi K3.1 Pro and K3.1 Pro Highspeed through its web products and API. The base model accepts text and images, offers a 1-million-token context window and is priced at $2 per million input tokens, $10 per million output tokens and $0.20 per million cached-input tokens. Moonshot says Highspeed is two to three times faster at roughly twice the regular input/output rates. The company calls K3.1 Pro open-weight in its announcement but explicitly says the full weights will only be released by November 4; as of October 25 they are not yet available. Independent Artificial Analysis testing gives K3.1 Pro an Intelligence Index score of 55, 67% on Terminal-Bench 4.0, and an average cost around $2.11 per benchmark task.
Why it matters
A frontier-class coding and agent model available at a lower API rate than many leading alternatives changes model-selection economics for builders who measure outcomes rather than only tokens. But the model is verbose, and a cheap token price does not guarantee a cheap completed task. The independent benchmark's 73.5 output tokens/second is below its 81-token/second comparison median, and its task cost and long-running performance matter for agent workloads. Local deployment, licence terms and unrestricted fine-tuning cannot be assumed until the weights and licence actually ship.
The independent comparison is promising but task-specific
Artificial Analysis ranks K3.1 Pro fourth on its October 23 Intelligence Index with a score of 55, and reports 67% on Terminal-Bench 4.0 versus 13% for Kimi K3. Its test records 55,000 output tokens per evaluated task and roughly $2.11 per task. The vendor separately reports 65.7% on Terminal-Bench 4.0 under its own setup, a reminder that harness, settings and task sampling can change scores. Simon Willison independently tested API availability with a small visual coding prompt, not a full production reliability audit.
What the pricing actually buys
The published Kimi API price is $2 input and $10 output per million tokens, with cached input at $0.20. The optional Highspeed endpoint is advertised as two to three times faster but more expensive; teams should benchmark end-to-end time and successful-task cost rather than assume it always wins. A 1M context window and text-plus-image input broaden possible workloads but do not establish reliable recall at maximum length.
Open weights are still a future milestone
Moonshot promises to release full K3.1 Pro weights by November 4, 2026. Until downloadable files, a licence and deployment guidance are actually available, K3.1 Pro should be treated as a proprietary hosted model for procurement and compliance decisions. Moonshot says it deliberately removed some cyber-related training data for the planned open release; its safety measurements are vendor-produced and should not be confused with independent security validation.