Kimi K3.1 Pro goes live on Moonshot API: weights promised by November 4, Highspeed 2–3x faster with same cache-hit pricing
According to a linked post by Simon Willison on October 23, Kimi K3.1 Pro went live on the Moonshot API that day, with weights not yet open; the official promise is to release the full weights by November 4 at the latest. The Highspeed version is 2–3x faster at double the price, with the same price for input cache hits. The author ran a single SVG test via the OpenAI-compatible endpoint and compared the three reasoning-effort levels of the two model generations.

What changed
According to a linked post by Simon Willison on October 23, 2026, Kimi K3.1 Pro landed on Moonshot's Kimi API that day under the model name kimi-k3.1-pro; the version is not yet open-weight, and Moonshot promises to release the full model weights by November 4, 2026 at the latest. The same model also has a Highspeed version, about 2–3x faster at double the price, but input-token cache hits are priced the same as the standard version. The USD prices cited in the post are $2 per million input tokens and $10 per million output tokens, which the author says matches GPT-6.1 Sol and is about half of Claude Opus 5.5. On evaluation, Moonshot's own benchmarks show Kimi roughly on par with Opus 5.5 and GPT-6 Astra, leading on a few benchmarks and close on others; Artificial Analysis gives it 55 points, above Claude Fable 5.1, GPT-6 Astra and Gemini 4 Argon at 53, and below Claude Opus 5.5 at 58. On that basis the author argues this is the first time a (soon-to-be) open-weight model has scored this high in a frontier comparison. He called the model with a command-line tool that supports OpenAI-compatible endpoints, generated an SVG from a single prompt, and noted that the tool does not log results by default and reasoning traces must be redirected and saved separately; he then compared K3.1 Pro and Kimi K3 at low, medium and high reasoning effort.
What it means for developers
For API users, the practical changes in this version come down to three points. On access, the model uses Moonshot's OpenAI-compatible endpoint, so existing OpenAI-compatible clients can largely try it by switching the model name, but the calling tool mentioned in the source post does not persist output by default and reasoning traces need to be redirected and saved, otherwise comparisons across reasoning levels lack raw material to refer to. On version tiers, Highspeed and the standard version are not a simple speed-for-money trade: input tokens that hit the cache are priced the same, meaning that in scenarios with fixed prefixes and long system prompts that hit the cache often, the extra cost paid for speed may be well below the apparent doubling; while for batch tasks with low cache-hit rates and a high share of output tokens, the cost difference is close to double. On selection reference, the vendor-reported benchmarks only show rough parity with Opus 5.5 and GPT-6 Astra, and the third-party intelligence index is still below Opus 5.5, so treating K3.1 Pro as a peer replacement lacks public evidence; meanwhile the weights are not out yet, so any self-hosting or private evaluation must wait until after the promised date, and with license terms unknown, the self-hosting cost calculation cannot be done yet.
Integration and evaluation
The following are conditional suggestions based on the version tiers and calling method described in the source post, not verified conclusions. First, how to choose between the standard and Highspeed versions: first measure your requests' input cache-hit rate and output-token share, then run the same batch of prompts through both versions and record the total cost per completed task and p95 latency; Highspeed may pay off only when the cache-hit rate is high enough and the task is latency-sensitive, while offline batch processing with low hit rates should prefer the standard version. Second, how to set reasoning effort: the source post mentions low, medium and max levels; first split tasks into short structured extraction and long-chain agent categories, then measure task success rate and token consumption level by level on a fixed held-out set, and set the default level from the relationship between success rate and cost rather than copying the result of a single SVG example. Third, whether to wait for the weights: if you have requirements for data not leaving your premises, need fine-tuning, or have very high monthly usage, treat November 4 as a checkpoint and wait until the weights and license terms are public before calculating self-hosting costs; until then, small-traffic comparisons via the API are the lower-cost approach.
Limits and open questions
kimi-k3.1-pro does not currently appear in KimiSeek's live catalog; the API availability and pricing described here refer to Moonshot's official channel, and whether other platforms offer it and how they price it needs separate confirmation; the live catalog has a Highspeed entry in the same family, but that does not mean the standard version is live on that platform. Open weights are not the same as an open-source license: even if the full weights are released by the promised date, the license scope and commercial-use and redistribution terms have not been publicly explained, so one cannot infer free use or fine-tuning from it. Prices change; the $2 and $10 figures are the prices recorded in the source post on October 23, 2026, and before integrating, refer to the vendor's or the platform's pricing page. The strength of the evidence is also limited: Moonshot's benchmarks are vendor-reported, the Artificial Analysis score and the claim of the "first high-scoring open-weight model" come from the author's account, and the SVG generation and three-level reasoning comparison are results from a single prompt and do not represent general capability; actual selection still needs verification on your own task set.
Illustrations from the original report
