Kimi K3.1 Pro is on OpenRouter, with 1M context and weights due by Nov 4

Moonshot AI’s Kimi K3.1 Pro has appeared on OpenRouter under the ID moonshotai/kimi-k3.1-pro. According to Moonshot’s own documentation, it is a 1M-token model whose full weights are due by November 4.

The OpenRouter page is mostly an API usage guide. It lists no description, price or context length in the text we could read, so most of what follows comes from Moonshot’s documentation and its launch blog post, plus third-party listings. Benchmark results are the company’s own.

A post-trained K3 with a 1M-token window

Moonshot describes K3.1 Pro as the first model in a new K3.1 series and a post-trained version of Kimi K3 with substantially more post-training compute. It calls it its most capable “3T-class” open-weight model so far. The context window is 1,048,576 tokens, and the maximum output length is the same, according to the Vercel AI Gateway listing.

Thinking mode is always on and cannot be switched off. Requests that try to disable it return an error, the documentation says. The reasoning_effort setting takes low, high or max, and max is the default. Text, image and video can go in, and text comes out. Images must be sent as base64 data URLs or uploaded file references, not public links. Structured output, custom tools and automatic context caching are supported.

A faster sibling, kimi-k3.1-pro-highspeed, is described as running two to three times faster. On OpenRouter, the page says in-region routing for the EU is not currently available for this model.

Cheaper per token than the closed leaders, if the numbers hold

Neither OpenRouter’s page nor Moonshot’s quickstart prints a dollar price. The Vercel AI Gateway lists K3.1 Pro at $2 per million input tokens and $10 per million output tokens. OpenRouter’s own rate may differ, so check the model page before budgeting. Moonshot says its API pricing is flat, with no extra tier for long contexts, and that cache hits are billed at a lower rate than misses.

The flat rate matters more than usual at this window size. A request that fills the 1M-token window at the Vercel input price would cost about $2 before any output, and reasoning tokens on the max setting are billed as output. Moonshot sets the default output cap at 131,072 tokens, so a single maxed-out call could add roughly $1.31 on top at the listed output rate. Developers who use the model for agents should therefore watch the effort setting as closely as the headline price. Our API pricing tracker has the per-million-token costs of other major models for comparison.

On performance, Moonshot’s blog puts K3.1 Pro at 65.7% on Terminal-Bench 4.0. It lists GPT-6 Astra at 58.2% and Claude Opus 5.5 at 66.4% on the same test. On the BenchCAD 3D-modeling test with tools, it scored 92.7, behind Claude Opus 5.5 at 96.2 and GPT-6 Astra at 95.9. These figures come from Moonshot’s evaluations, not independent tests. Moonshot’s quickstart also still shows unfilled placeholders for some scores, and a figure for the company’s hallucination test differs between its pages and press coverage, so treat the numbers as provisional.

What is still unproven about the open weights

Moonshot says the full weights will be released by November 4. The documentation does not say whether that covers the Highspeed variant, and it names no license. Until the files are published, K3.1 Pro can only be used through hosted APIs, which is the situation that makes the OpenRouter listing useful: developers can test it with one API key and compare it with other models.

The company also said it removed cyber-related data from mid-training and reinforcement learning. On its internal xdev-bench-hard test, a checkpoint without those measures scored 90.5% and the final model 81.0%. Moonshot adds that the model’s cyber capability remains behind frontier proprietary models that have been public for at least three months. The measure is the company’s own, and whether such data removal survives fine-tuning of open weights is something outside testers will need to check once the files are out.

Moonshot has had API turbulence before; see our look at Kimi K2 Thinking and what happened to its API. On Moonshot’s own platform, access requires a top-up of at least $1, and rate limits rise with cumulative spend. Unite.AI reports that on Kimi’s consumer plans the model requires the Plus tier or above, with the full 1M context reserved for Pro and above.