SI-APIPortuguês (BR)Open console MODEL RELEASE
Kimi K3.1 Pro is here: API pricing, benchmarks and when to use Highspeed
Moonshot has launched Kimi K3.1 Pro with a 1M-token context window and standard API rates of $2 for input and $10 for output per million tokens. Highspeed promises faster output at higher base rates. Hosted access is available now; full model weights are scheduled for release by November 4.
What changed from Kimi K3?
K3.1 Pro builds on K3 with more post-training work. Moonshot's examples focus on coding agents that work through longer tasks, scripted 3D modeling and financial documents. The company describes it as a 3T-class model.
For access today, use Kimi's app, Kimi Work, Kimi Code or the official API. The weight release is a separate milestone: Moonshot says the full weights will arrive by November 4, 2026.
How much does the official API cost?
Pro costs $2 per million input tokens and $10 per million output tokens. Highspeed doubles those base rates, while cached input costs $0.20 for either model.
| Model / API ID | Input | Output | Cache hit |
|---|---|---|---|
Prokimi-k3.1-pro | $2.00 | $10.00 | $0.20 |
Pro Highspeedkimi-k3.1-pro-highspeed | $4.00 | $20.00 | $0.20 |
Cache writes cost extra: $2 / $4 per million tokens for Pro with a 5-minute / 1-hour cache lifetime, or $4 / $8 for Highspeed. The default lifetime is 5 minutes.
A simple cost estimate: 1 million ordinary input tokens plus 200,000 output tokens comes to $4 on Pro or $8 on Highspeed, before cache writes and taxes. This is arithmetic using the listed rates, not a measured agent run.
Kimi bills API usage separately from Kimi membership and Kimi Code subscriptions. Check the current price sheet when budgeting a larger workload.
What do the launch benchmarks show?
Moonshot reports 65.7% on Terminal-Bench 4.0 and 61.2% on its internal Kimi Code Bench 3.0 Long-horizon, up from K3's 55.95% on the latter test.
| Evaluation | K3.1 Pro | What it measures |
|---|---|---|
| Terminal-Bench 4.0 | 65.7% | Agents completing terminal tasks under test constraints. |
| Kimi Code Bench 3.0 Long-horizon | 61.2% | Multi-step engineering work in Moonshot's internal evaluation. |
These are company-reported results. Moonshot used the Claude Code harness for both evaluations and generally set reasoning effort to max. The agent setup matters when comparing scores. We have not reproduced these runs.
What is worth trying beyond chat?
Moonshot demonstrates projects in game engines, 3D assets built with modeling tools and financial workbooks. Treat the demos as examples of the intended workflows. For your own trial, pick an existing task whose result you can inspect: a repository fix with tests, a model with specified dimensions or a spreadsheet with traceable formulas.
A large context window lets you supply more material. It still pays to check whether the output uses the right files and cites the right evidence.
Should you pay for Highspeed?
Start with Pro for background work. Consider Highspeed when you spend enough time waiting for generated output to justify the higher rates. Moonshot advertises 2 to 3 times faster output; that does not establish the same speedup for a whole task that also uses external tools.
Pro: a sensible first trial for batch jobs and tasks you can leave running.
Highspeed: worth testing for interactive coding and other work where waiting interrupts you.
Run the same task on both versions with matching settings. Compare completion quality, total elapsed time and the bill, including cache usage. Choose based on that result.
How do you try it?
In Kimi Code, use /model to select K3.1 Pro. For the API, use the model IDs above. The official guide lists low, high and max reasoning effort, with max as the default; thinking stays enabled.