Moonshot AI launched Kimi K3.1 Pro on October 23, 2026 as the first model in the K3.1 series: a post-training of Kimi K3 at the 3 trillion parameter scale with 1 million tokens of context, backed by a reworked reinforcement learning stack covering hundreds of downstream domains. A Highspeed variant ships the same intelligence at roughly two to three times the output speed for twice the price. Vendor-reported figures put K3.1 Pro at 65.7% on Terminal-Bench 4.0 (Claude Opus 5.5: 66.4%, GPT-6 Astra: 58.2%), 92.7% tool-assisted on BenchCAD, and 64.4 on OfficeQA Pro v2, listed as the best published result; on the internal Kimi Code Bench 3.0 Long-horizon it scores 61.2% against K3's 55.95%, and the Kimi Hallu Bench Hard stress-test single-fact error rate falls from 23.4% to 8.3%. Launch demos cover Unreal Engine 5.8 and Godot 4.7.2 game builds, a FreeCAD turbofan with 2,100 solids, Blender scenes, and a walking simulator where the model debugged 36 errors in its own simulator's shoulder muscle moment-arm calculation. The API on platform.kimi.com lists kimi-k3.1-pro at ¥1.3 per million cached input tokens, ¥13 uncached, and ¥68 per million output tokens, with the Highspeed variant at twice those rates. Full model weights are promised before November 4, 2026, after attack-and-defense data was removed from both the mid-training and reinforcement learning stages; on the vendor's xdev-bench-hard the unfiltered model would have solved 90.5% on pass@3 and the shipping weights solve 81.0%. All figures are from the launch post and have not been re-run.
Moonshot AI launched Kimi K3.1 Pro on October 23, 2026, the first model in the K3.1 series and its strongest to date at the 3 trillion parameter scale, with 1 million tokens of context. It is a post-training of Kimi K3 with much more post-training compute, and Moonshot reports its largest gains in coding, 3D creation, financial research, and long-horizon agent work. On Terminal-Bench 4.0 it scores 65.7%, above GPT-6 Astra's 58.2% and just under Claude Opus 5.5's 66.4%, per the launch post. A Highspeed variant runs at roughly two to three times the output speed for twice the price. You can use both today on kimi.com, in the Kimi app, and through the API, where kimi-k3.1-pro costs ¥1.3 per million cached input tokens, ¥13 uncached, and ¥68 per million output tokens. Full model weights are promised before November 4, 2026.
Want to try the latest Kimi K3.1 Pro and Kimi K3.1 Pro Highspeed? Explore the SaveMyToken Multi-Model Coding Plan.
What is Kimi K3.1 Pro?
It is the first release in the K3.1 series, and technically a post-training of Kimi K3. The base model is unchanged in identity; what changed is the amount of post-training compute behind it. To spend that compute across hundreds of downstream domains at once, Moonshot reworked the whole stack: the RL framework, long-context inference, training efficiency, sandbox infrastructure, the scheduling system, and a shared platform for checking environment quality. The company's summary is that every domain picked up large-scale RL gains rather than a few areas advancing at the expense of the rest.
Two models shipped. Kimi K3.1 Pro is the standard one. Kimi K3.1 Pro Highspeed has the same intelligence with roughly two to three times the output speed, according to the launch post, at twice the price. Full model weights for the open release are due before November 4, 2026.
How does it score against GPT-6 Astra and Claude Opus 5.5?
The headline numbers put K3.1 Pro in the same band as the strongest closed models, with the usual caveat that every figure below comes from the vendor's launch post and none have been re-run here.
| Benchmark | Kimi K3.1 Pro | Kimi K3 | Claude Opus 5.5 | GPT-6 Astra |
|---|---|---|---|---|
| Terminal-Bench 4.0 | 65.7% | not disclosed | 66.4% | 58.2% |
| BenchCAD, tool assisted | 92.7% | not disclosed | 96.2% | 95.9% |
| OfficeQA Pro v2 | 64.4, listed as best score | not disclosed | not disclosed | not disclosed |
| Kimi Code Bench 3.0 Long-horizon (internal) | 61.2% | 55.95% | not disclosed | not disclosed |
| Kimi Hallu Bench Hard, single-fact error rate (internal, lower is better) | 8.3% | 23.4% | not disclosed | not disclosed |
All figures as reported in Moonshot AI's launch post of October 23, 2026. Terminal-Bench 4.0, BenchCAD, and OfficeQA Pro v2 are external benchmarks; the two Kimi-named tests are the vendor's own. Where the post gives no figure, the cell says so.
Terminal-Bench 4.0 measures coding agents on resource-constrained terminal tasks in a real environment, with a strict dedicated test attached to each task. The 65.7% puts K3.1 Pro within 0.7 points of Claude Opus 5.5 and 7.5 points above GPT-6 Astra. BenchCAD, where the model writes CAD code to rebuild a 3D object from multi-view renders, ends with K3.1 Pro at 92.7% against 96.2% and 95.9% for the two closed leaders, close but clearly behind on this one.
The fact-error number deserves its own sentence. Kimi Hallu Bench Hard is an internal stress test that deliberately concentrates on the cases where models tend to slip, so the rates are not representative of everyday use. On that hard subset, the share of single-fact statements that were wrong fell from 23.4% in K3 to 8.3%, and the share of answers containing at least one factual error dropped to about 42.4% of the previous level. A stress test improving by that much is the more meaningful signal, because that is exactly where a finance or research workflow lives or dies.
What can it do with code?
Moonshot's range for the model runs from fixing a bug in an existing codebase to building a complete project from zero, and the launch demos lean hard into the second half. Three shipped end to end from single prompts, per the post. One built a celestial-palace battle game in Unreal Engine 5.8: a palace on a sea of clouds with a 1,750 meter colonnade of 36 bays, procedurally modeled in Blender, 36 million source triangles rendered through Nanite and Lumen, with cloud volumes, 8K materials, and NASA lunar elevation data pulled in as external assets. A second built a stylized 3D action-adventure game in Godot 4.7.2 with five explorable regions, third-person combat, a two-phase boss fight, gathering, a day-night cycle, and campfire fast travel. The third was a browser page showing a 3D robotic arm with exploded views and a transparent shell, its 2,317 parts and 268 animated wire routes all generated in code.
The long-horizon claim rests on a messier internal demo, and it is the more convincing one. Given an open-source human anatomy model with 296 skeletal components and 402 muscles, K3.1 Pro built a walking simulator with its own training and evaluation pipeline, then iterated through 167 tasks and 420 independent reviews. The simulated body went from collapsing within its first steps to walking continuously for 11.2 seconds. When reward tuning failed to fix the arm swing, the model probed the simulator's dynamics open-loop, found 36 errors in the shoulder muscle moment-arm calculation, corrected them, and retrained the controller. Debugging the simulator rather than the policy is the part a benchmark score does not show.
On the vendor's internal Kimi Code Bench 3.0 Long-horizon, which tests multi-step execution and tool use across engineering domains, K3.1 Pro scores 61.2% against K3's 55.95%.
What about 3D work?
3D creation gets its own section in the announcement, and the tool-assisted BenchCAD score of 92.7% is the external anchor for it. In practice the model writes and runs scripts for Blender, FreeCAD, and other modeling tools directly, without going through MCP connectors, which keeps parametric work like mechanical parts, building models, configurable products, and custom furniture inside one conversation.
The demos are specific enough to check against the tools named. A FreeCAD turbofan assembled in Python across 2,100 solids, with staged assembly, exploded views, and rotation animation at 2,780 RPM for the fan, 8,590 for the low-pressure rotor, and 12,400 for the high-pressure rotor. A Blender scene of an airport indoor rainforest valley rendered to a 17.5-second clip at 1920×1080 and 60 fps, citing its Poly Haven HDRI by name and license. A five-part 3D-printable desk lamp designed with trimesh, styled after the round form of Fujian tulou buildings.
A 3D Make feature shipped on kimi.com and the Kimi app alongside the model. It previews and edits 3D models in the browser, sends finished models to download for printing, and exports video or live photos for sharing.
Can it carry finance work?
On OfficeQA Pro v2, an external benchmark built on real financial documents that requires retrieval, cross-document checks, and multi-step numerical reasoning, K3.1 Pro scores 64.4, which Moonshot lists as the best published result. On Frontier Finance, which covers full investment workflows from screening to profitability analysis, the post says performance improved over K3 without giving a figure.
Three internal benchmarks fill in the comparison against K3: up 14% on Financial Modelling, which grades formula linkage, financial logic, and workbook quality; up 20% on Financial Research, which requires verifiable output under evidence and deadline constraints; and 4.2 times the K3 score on Business Operation, a simulated long-horizon management environment. The demos include an end-to-end NVIDIA financial model with linked formulas and adjustable drivers, a CATL investment-committee deck organized around expectation gaps against consensus, and a quantum-computing industry report whose citations the post says trace to primary sources.
The Hallu Bench Hard result in the table above is the number to watch for this category, and Moonshot states the caveat honestly: the test deliberately samples error-prone cases, so its rates overstate everyday error levels.
What does the API cost?
| Model ID | Input, cache hit | Input, cache miss | Output |
|---|---|---|---|
| kimi-k3.1-pro | ¥1.3 / M tokens | ¥13 / M tokens | ¥68 / M tokens |
| kimi-k3.1-pro-highspeed | 2× standard rate | 2× standard rate | 2× standard rate |
List prices on platform.kimi.com in RMB, as stated in the launch post. The Highspeed row is the post's stated 2× multiplier; applied to the standard rates it works out to ¥2.6, ¥26, and ¥136 per million tokens, though no separate listing was shown in the announcement.
Two things in that table are worth more than the headline rates. First, a cache hit costs exactly one tenth of a cache miss, so workloads with stable prefixes, long system prompts, or repeated document context get an order-of-magnitude discount on the input side. Second, Highspeed is priced at twice the rate for two to three times the speed, which only pays off when latency feeds back into cost or user patience. An interactive agent that burns tokens while a user watches benefits; an overnight batch job gains nothing and pays double.
When do the open weights land, and what got removed?
Full model weights are due before November 4, 2026. The interesting part of the announcement is what Moonshot says was taken out before that release.
During development, the team watched vulnerability-research performance climb as post-training scaled, and traced the source: attack-and-defense content had entered the mid-training corpus and the reinforcement learning environments as a byproduct of training on broad data. In preliminary internal experiments, leaving that data in across both stages pushed the downstream attack capability up sharply. For the open-weights build, the company removed it from both stages while keeping general coding tasks.
The before-and-after comes from the vendor's own xdev-bench-hard, an expert-curated test of source-based vulnerability discovery and exploitation built on recently disclosed flaws in software including the Linux kernel, Node.js, JavaScriptCore, Exim, and VirGL, with mitigations enabled. On pass@3, the unfiltered model would have solved 90.5%; the shipping weights solve 81.0%. Moonshot also states that K3.1 Pro's offensive security capability still trails closed frontier models that have been public for at least three months, and frames the removal as the first instance of an ongoing per-stage control policy.
The stated reasoning for releasing at all runs as follows: the weights do not advance the absolute frontier of attack capability; the open share of that capability will rise progressively rather than all at once; and defenders who need dual-use tools do better running strong models on infrastructure they control than leaving capability concentrated in a few closed systems. Whether that argument holds is a policy debate. What the announcement does is state the numbers that let you have it.
Where can you use it?
Both models are live now. kimi.com serves them in the browser, and the Kimi app covers iOS, Android, and HarmonyOS. Kimi Work, the desktop client for agentic work, runs on Windows and Apple-silicon Macs. Kimi Code runs in your terminal and desktop, with the model selectable as kimi-k3.1-pro or kimi-k3.1-pro-highspeed. The API is on platform.kimi.com at the prices in the table above.
For companies, Kimi's enterprise plan adds data isolation between personal and business accounts and member management, subscribable through the enterprise tab at kimi.com/membership/pricing. Kimi Hosted Agents offers the agent harness as a cloud service with long-horizon and multi-agent support, documented at platform.kimi.com/docs/hosted-agents/overview. A partner program is recruiting integrators to staff forward-deployed engineer roles.
The complete benchmark suite the announcement summarizes lives at kimi.com/blog/kimi-k3-1-pro.
Frequently asked questions
Is Kimi K3.1 Pro open weights?
Not on day one. The launch post commits to releasing the full model weights before November 4, 2026, and says attack-and-defense security data was removed from both the mid-training and reinforcement learning stages of the open-weights build. The license had not been named in the announcement at the time of writing.
How much does Kimi K3.1 Pro cost on the API?
On platform.kimi.com, kimi-k3.1-pro lists at ¥1.3 per million input tokens on a cache hit, ¥13 per million on a cache miss, and ¥68 per million output tokens. The Highspeed variant, kimi-k3.1-pro-highspeed, costs twice those rates for roughly two to three times the output speed.
Should you pay twice for Kimi K3.1 Pro Highspeed?
Pay the 2× rate when latency feeds back into cost or user patience, such as interactive agents that burn tokens while a user waits. For offline batch generation where speed does not compound, the standard model produces the same output for half the price.
How is Kimi K3.1 Pro different from Kimi K3?
K3.1 Pro is a post-training of K3 with substantially more post-training compute, applied across hundreds of downstream domains through a reworked RL stack. On Moonshot's internal Kimi Code Bench 3.0 Long-horizon it scores 61.2% against K3's 55.95%, and on the Kimi Hallu Bench Hard stress test its single-fact error rate falls from 23.4% to 8.3%.