Moonshot released the Kimi K3.1 Pro weights just after midnight on 5 November, Beijing time. The repository is on Hugging Face. The model is a 3T-class open-weight model with a 1,048,576-token context, native visual understanding and tool calling.
The price moved more than the benchmark table did.
Official list price is $2.00 per million input tokens, $10.00 per million output tokens and $0.20 per million cached input tokens. K3 lists at $3.00, $15.00 and $0.30. That is one-third lower on all three lines.
The output line is the bill. A model that finishes a task in fewer turns can be cheaper than a model with a lower input rate and a longer loop, because at $10 per million output tokens the writing is what you pay for.
What Moonshot says changed
Moonshot describes K3.1 Pro as a post-trained version of K3 with substantially more post-training compute. The release notes claim consistent gains across coding, 3D creation, financial research and agentic work.
The published numbers:
| Benchmark | K3.1 Pro | Comparison |
|---|---|---|
| Terminal-Bench 4.0 | 65.7% | Claude Opus 5.5 66.4%, GPT-6 Astra 58.2% |
| Kimi Code Bench 3.0, long-horizon | 61.2% | K3 55.95% |
| BenchCAD, with tools | 92.7% | Claude Opus 5.5 96.2%, GPT-6 Astra 95.9% |
These are Moonshot's numbers. We did not run Terminal-Bench or BenchCAD, so treat them as the vendor's claim until an independent run exists.
The model card also lists a highspeed variant at 2-3x the speed of the standard model. Official price for that variant is $4.00 input, $20.00 output and $0.20 cached.
What we measured before the release
On 2 November, three days before the release, we ran 5,315 Japanese, German and Turkish tasks against the K3.1 Pro API. The run returned 5,368 responses and 1,551,208 tokens.
| Task | Result |
|---|---|
| Japanese idiom versus literal, OpenMWE | 270/297, 90.9% |
| Turkish paraphrase selection | 199/200, 99.5% |
| Turkish figurative versus literal | 195/197, 99.0% |
| German conversational meaning, MultiPragEval | 263/300, 87.7% |
| German indirect answers, InQA+ | 256/384, 66.7% |
| Turkish conversational implication, ImplicaTR | 2,179/3,888, 56.0% |
The full rows, the prompts and the scoring sheets are downloadable from the test page. If you want to argue with a number, start there.
What we are checking before we open
Three things can move a hosted score after release day.
The license on the model card. We host models only where the license allows commercial serving. The card governs that, not the benchmark post.
Weight parity. The API we tested in November is not the same artifact as the released weights. Quantization and the serving stack can move scores by a few points. We will re-run a subset against our own deployment and publish both numbers.
Throughput on our path. We have not published a time-to-first-token or sustained tokens-per-second number for K3.1 Pro on the Wallaby endpoint. When we do, it will carry the network, the sample size and the date.
Until those checks pass, the status is simple. The model is available from Moonshot today. Our endpoint opens shortly at $1.80 input, $9.00 output and $0.18 cached input per million tokens, 10% below the official list.
Why the price cut matters more than the benchmark
A 5.25-point gain on a coding benchmark is useful. A one-third cut on the output rate changes which workloads are worth running. Long-horizon coding, document passes over a 1M-token context and multilingual review all spend most of their money on output tokens.
Open weights do not make inference free. They make it contestable: you can self-host, or buy the same weights from a provider that meters them. We are in the second business, and our rate card is public.
Our endpoint opens shortly. The 5,315-task test is already up.