Kimi K3.1 Pro Pricing, Access and Open Weights Release Date
Kimi K3.1 Pro costs half of Opus per token on the Kimi API, with open weights promised by November 4. Specs, access limits and what is still unconfirmed.
How much does Kimi K3.1 Pro cost and when do the open weights ship?
Kimi K3.1 Pro is live now on the Kimi API at half the per-token price of Claude Opus, according to coverage of the launch. Moonshot AI has promised the open weights by November 4. The model is a mixture-of-experts system with about 2.8 trillion total parameters, roughly 104 billion active per token, and a 1 million token context window. The sources give no absolute dollar price per million tokens, only the comparison to Opus.
TL;DR
- Kimi K3.1 Pro is available today on the Kimi API at half Opus's per-token price, though no exact per-million-token figure appears in the sources.
- The open weights are promised by November 4. Until they ship, that date is an announced plan, not a release.
- The model has 2.8 trillion total parameters and about 104 billion active, the same size as Kimi K3, with a 1 million token context window.
- A high-speed variant runs two to three times faster than the standard model.
- On Moonshot's cost chart, the model reaches Opus-class scores at about a third of Opus's cost per task, though one reviewer put Opus at roughly three times the cost on one benchmark.
- Moonshot removed cyber-related data from training, so the open weights will be weaker at offensive security than its internal checkpoint.
What is Kimi K3.1 Pro?
Kimi K3.1 Pro is a post-trained version of Kimi K3 from Moonshot AI. It keeps K3's size and adds a lot more reinforcement learning. Moonshot ran large-scale RL across hundreds of domains and rebuilt its training stack so that extra RL compute keeps producing steady gains.
The model targets coding, 3D creation, financial research and long agentic tasks. Moonshot claims it matches or comes close to the best closed models.
| Spec | Reported value |
|---|---|
| Architecture | Mixture of experts |
| Total parameters | 2.8 trillion |
| Active parameters | About 104 billion |
| Context window | 1 million tokens |
| Relationship to K3 | Same size, more post-training RL |
| Fast variant | 2 to 3 times faster |
One reviewer described it as a "3 trillion parameter class" model. The 2.8 trillion figure from the other source is more specific, so use that one.
How much does Kimi K3.1 Pro cost?
The headline number is relative: the Kimi API charges half of what Opus charges per token. The sources do not state the Opus price or the Kimi price in dollars, so check the current Kimi API pricing page before budgeting. For a dollar breakdown, see Kimi K3.1 Pro API pricing and usage limits.
Per-token price is only part of the cost. What you pay per finished task also depends on how many tokens the model burns. Two figures from the launch coverage point the same way:
- On Moonshot's own chart, the model reaches its Terminal Bench 4.0 result for about a third of what Opus costs per task.
- In one reviewer's reading of the Terminal Bench 4 cost chart, K3.1 Pro scores about 66% at roughly $5 per task. Claude Opus 5.5 scores about the same but costs close to three times more. Sonnet 5.5 costs around four times more for a lower score.
These are vendor-supplied and reviewer-read numbers, not independent audits. Treat the ratios as directional.
How can I use Kimi K3.1 Pro right now?
There are three routes mentioned in the sources.
- The Kimi API. The model is live there today, and this is the route to use for production testing and for plugging into your own tooling.
- The Kimi app and harness. One creator downloaded the Kimi app and its harness and ran K3.1 Pro in a "High" configuration. Another configured a Kimi coding harness to use K3.1 Pro and pointed it at a local Docker project.
- The model picker for the fast variant. A high-speed version exists and runs two to three times faster. The sources do not say how it is priced, so confirm that before switching.
Expect friction. During launch-day testing, the model was described as open to everyone, but an account or a paid subscription appeared to be required because traffic was so high. That is a snapshot from one livestream, so it may change. Moonshot also closed off subscriptions for a while a few months earlier because of demand, which suggests capacity limits are a real pattern with this lab.
When will the Kimi K3.1 Pro open weights be released?
Moonshot says the weights will arrive by November 4. Both launch write-ups repeat that date. The wording matters: "by" a date and "promised" mean the release is planned, not shipped. Nothing in the sources confirms a license, a download location, quantized builds or hardware requirements.
Reviewers also said they expect fuller architectural details once the weights are public. Until then, the only architecture facts are the mixture-of-experts design, the parameter counts and the context length.
Can you run Kimi K3.1 Pro locally?
The sources do not say, so this is simple arithmetic rather than a reported fact. A 2.8 trillion parameter model needs well over a terabyte of memory just to hold the weights at 4-bit precision, before you count the KV cache for long contexts. That puts it in multi-node server territory, not a single workstation. The 104 billion active parameters reduce compute per token, but all the experts still need to be resident or paged in.
For most builders, the practical path is the hosted API now, and a third-party host or a self-managed cluster after the weights land.
How does Kimi K3.1 Pro perform?
Moonshot and the two reviewers report the following:
- Terminal Bench 4.0. Less than a point behind Claude Opus 5.5 and ahead of GPT-6 Astra, per one source. The other described it as neck and neck with Opus 5.5.
- Other leads. It leads on several benchmarks, including long coding tasks, the Agents Last Exam and Humanity's Last Exam with tools.
- Finance. It took first place on Office QA Pro V2, ahead of Opus 5.5. On Frontier Finance it placed second, behind Claude Fable 5.1 and GPT-6 Astra.
- Where it trails. It does not win everything. It is behind the leaders on a science-style terminal task and on one other benchmark, though not by a wide margin.
Hands-on tests showed a similar mix. The model built single-file HTML scenes (a pizza oven, a doner kebab broiler) and a 3D "floor is lava" game from one prompt. In a six-service Docker app, it found a hidden currency-conversion bug from a vague instruction, rebuilt only the affected API and tested against the live database. It produced a five-year, three-statement Excel model with a DCF and 343 recalculated formulas. It also handled a request for band names across 81 languages in a strict format.
Vision was weaker. In a seatbelt check on an AI-generated image, it wrongly said the shoulder strap lay correctly, which the reviewer flagged as an error. One test is thin evidence, but it is a reminder to verify image-based outputs.
Moonshot also shared a walking simulation of the human body. When the walk stopped improving, the model went back and checked the simulator itself. Moonshot says it found 36 errors in the shoulder muscle math.
What did Moonshot change for security?
Moonshot removed cyber-related data from both pre-training and RL before the open release. The stated reason is that the model's exploit skills kept climbing during training. As a result, the open weights have less offensive hacking skill than Moonshot's internal checkpoint.
Even so, Moonshot says the model solves more than twice as many exploit tasks as Kimi K3 did. If you evaluate it for security-adjacent work, test it directly instead of assuming the removal made it harmless.
Is Kimi K3.1 Pro worth trying now?
If you already run agentic coding or finance workloads on Opus-class models, a half-price API at similar benchmark results is worth a trial run. Use your own tasks, and measure tokens per completed task, not just the per-token rate. The reviewer observations about lower token use and fast bug fixes are encouraging but anecdotal.
If you need self-hosting, wait for November 4 and for the license and file details. If you depend on vision accuracy, test carefully first.
Frequently Asked Questions
Is Kimi K3.1 Pro open source yet?
Not yet. Moonshot has promised the open weights by November 4, and the model is available through the Kimi API in the meantime. Check that the weights and license actually appear before building a plan around self-hosting.
How big is Kimi K3.1 Pro?
It has about 2.8 trillion total parameters with roughly 104 billion active per token, and a 1 million token context window. It is the same size as Kimi K3, with more reinforcement learning on top.
Is there a faster version of Kimi K3.1 Pro?
Yes. A high-speed variant runs two to three times faster than the standard model. The sources do not give its price or any quality trade-offs.
Do I need a paid account to use Kimi K3.1 Pro?
During launch-day testing, an account or paid subscription appeared to be required because of high traffic. That may relax as demand settles. The API is the other route, and it is billed per token.
Will the open weights match the hosted model?
Not entirely. Moonshot removed cyber-related data from training, so the open weights have less offensive security skill than its internal checkpoint. The sources do not say whether other differences exist.
How we source, write, and correct our articles. Editorial standards
