On October 23, 2026, Moonshot AI launched the AI model Kimi K3.1 Pro and its faster variant, Highspeed. The independent evaluator Artificial Analysis raised the model's composite score from 44 for the previous Kimi K3 to 55, an 11-point gain, placing it 4th overall. However, the release of the trained weights as an "open-weight" model is still planned for no later than November 4. The improvement stands out most when the model works with external tools to run code and programs and tackles long, multi-step tasks, rather than simply answering questions. By contrast, gaps with top models remain in knowledge questions answered without tools and in tests that strictly judge how complete a deliverable is, so assessing the model's competitiveness requires separating task types and evaluation conditions.
55 on the independent index, up 11 points from K3
Kimi K3.1 Pro is a model in which post-training, the stage that strengthens specific capabilities after pretraining, was reinforced following K3's pretraining. In its official blog, Moonshot positions K3.1 Pro as the post-trained version of K3. To carry out reinforcement learning across hundreds of domains, it says it optimized not only the training and inference infrastructure but also the scheduling of processing and the quality control of isolated execution environments. It has not disclosed how much each individual improvement contributed to the performance gain.
According to Moonshot's official X post, Kimi K3.1 Pro has 2.8 trillion total parameters, of which 104 billion are active during inference. Its context length, the amount it can process at once, is 1 million tokens. A token is the basic unit an AI model uses to process text and other data. Still, the model's sheer scale should be distinguished from the factors that actually drove this performance gain.
In the results Artificial Analysis published on October 23, its "Intelligence Index," which measures overall performance, rose 11 points from K3's 44 to 55. Claude Opus 5.5 leads at 58, with Claude Sonnet 5.5 and Muse Wonder Full tied at 56, putting Kimi K3.1 Pro in 4th. Note that Opus and Sonnet were run with a "max with fallback" reasoning setting and Muse with "max," so conditions are not perfectly uniform across models.
The largest improvement came in "Terminal-Bench 4.0," which measures the ability to solve problems by operating a terminal. In Artificial Analysis's evaluation, the success rate jumped from K3's 13% to 67%, surpassing Claude Sonnet 5.5's 64%. On "GDPval-AA," which compares how well models handle work-related tasks, it recorded 1846 Elo, taking 3rd place behind Opus at 1866 and Muse at 1857. Elo is a rating based on head-to-head comparisons between models, not a percentage of correct answers.
"AutomationBench-AA," which measures automation tasks, also improved from 58% to 73%. On the other hand, gains were relatively small on "HLE," which covers difficult knowledge questions (47% to 52%), and "SciCode," a scientific programming test (59% to 61%). In other words, the 11-point rise in the composite score does not mean every capability grew equally. This update is most worth testing for uses that involve running complex, long-running work with external tools.
Comparison with rivals: strong with tools, but completeness remains a weak spot
In the evaluation table Moonshot published on October 23, it compares results on "HLE-Full," which measures the ability to answer hard questions, with and without external tools. The table below summarizes scores for four major models. Reasoning effort is generally "max," but "max with fallback" applies to the non-public competing models, so execution environments and fallback conditions do not fully match.
| HLE-Full condition | Kimi K3.1 Pro | Claude Fable 5.1 | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|---|---|
| With tools | 70.0% | 65.6% | 57.2% | 67.7% |
| Without tools | 51.1% | 57.6% | 53.6% | 59.8% |
With tools, Kimi K3.1 Pro scored 70.0%, the highest of the four models compared. Without tools, however, it managed only 51.1%, below the other three. The result shows that the ability to use external tools and the ability to answer on the model's own should be evaluated separately. That said, the gap between the two conditions cannot be taken as the exact boost from tool use, and the contribution of the training method cannot be isolated. It also uses different evaluation conditions from the 52% HLE figure Artificial Analysis published, so the two should not be treated as the same measurement.
In programming, too, model rankings change between tests that sustain long-running work and tests that strictly judge the completeness of the final deliverable. Moonshot's comparison of six models is below. The figures are scores reported for each test and cannot be compared directly across different tests. Reasoning effort is generally "max," with "max with fallback" for non-public competing models, and Claude Code was used for Kimi's code-related tests.
| Test | Kimi K3.1 Pro | Fable 5.1 | GPT-6 Astra | Opus 5.5 | DeepSeek-V4.1-Flash-1017 | GLM-5.3 |
|---|---|---|---|---|---|---|
| Terminal-Bench 4.0 | 65.7 | 57.9 | 58.2 | 66.4 | 42.3 | 41.8 |
| SWE Marathon v1.1 | 51.7 | 45.6 | 42.5 | 46.3 | 40.0 | 42.5 |
| DeepSWE v1.1 | 75.5 | 67.4 | 74.1 | 74.2 | 77.0 | 66.9 |
| ProgramBench (binary scoring) | 31.5 | 53.5 | 50.0 | 65.0 | 29.1 | 19.0 |
*Fable 5.1 refers to Claude Fable 5.1 and Opus 5.5 to Claude Opus 5.5.
On "SWE Marathon v1.1," Kimi K3.1 Pro scored 51.7, the highest among those compared. On "ProgramBench," however, it managed only 31.5, a large gap from Claude Opus 5.5's 65.0. ProgramBench's binary scoring gives a 1 if a task's achievement score is 95% or higher and a 0 otherwise, so partially completed deliverables do not earn normal partial credit. In other words, being strong at sustaining long work does not guarantee meeting strict completion criteria. On "DeepSWE v1.1" as well, DeepSeek's 77.0 tops Kimi's 75.5, so Kimi cannot be said to lead in programming across the board.
It is also worth noting that this table is not an independent measurement of all models in the same environment. Competitors' Terminal-Bench scores use the maximum values posted on each company's blog or the official leaderboard, and DeepSWE uses figures from technical reports. For SWE Marathon, official leaderboard results are mixed with some of Moonshot's own evaluations, and execution environments, including the tools and procedures used, are not unified. Also, the 67% Artificial Analysis obtained on Terminal-Bench and Moonshot's reported 65.7% are separate evaluation results, and there is no basis for reading the difference as a change in model performance.
Therefore, simple win-loss judgments from narrow score differences should be avoided. When considering adoption, it is important to be clear about whether the target work requires long tool-driven operation, values the standalone accuracy of the model's answers, or needs deliverables that meet strict completion criteria. Even the same model can be rated very differently depending on the capability required.
Gains in 3D creation and financial document analysis
As practical applications of Kimi K3.1 Pro, Moonshot cites 3D creation and financial document analysis. On "BenchCAD," which measures the ability to generate CAD code from images shot or rendered from multiple viewpoints and reconstruct 3D shapes, Kimi scored 92.7%. That slightly exceeds Claude Fable 5.1's 92.6% but falls short of GPT-6 Astra's 95.9% and Claude Opus 5.5's 96.2%, while topping DeepSeek's 81.6%. In this test, 1,000 items were sampled from the full 17,900-item "Vision2Code" dataset, and publicly available models were evaluated using Claude Code. For non-public competing models, the companies' officially reported figures were used.
Moonshot showcases the AI's ability to generate and run scripts for Blender and CAD, and it also added a "3D Make" feature to the web and app versions. Users can preview and edit the generated 3D models and download them for 3D printing. However, a high BenchCAD score does not mean the output satisfies the dimensional accuracy required on manufacturing floors, or the certification and verification needed for business use. Being usable for prototyping shapes and exploring ideas is a separate matter from completing drawings that can be handed directly to a manufacturing process.
In document analysis, Kimi scored 64.4% on "OfficeQA Pro v2," exceeding Claude Opus 5.5's 62.2%. The other models scored 54.5% for Fable, 54.4% for Astra, 40.4% for DeepSeek, and 31.1% for GLM. In this test, entire PDFs are given to the model as images, with no machine-readable text data provided, and any answer that consults online sources scores zero. Because the execution environments were Codex for Astra and Claude Code for the other models, the scores cannot be read directly as differences in the models alone.
Moonshot also says its internal evaluations confirmed improvements in finance-related abilities. Compared with the old K3, it reports gains of 7.2 points in building financial models and 11.6 points in financial research. These are comparisons with its own previous model, however, and do not show an advantage over other companies' models. They are best treated as material for checking how far practical usefulness has improved in work that involves reading documents, performing needed calculations, and compiling analysis results into deliverables.
Higher performance also does not remove the need to fact-check generated content. In Moonshot's internal "Hallu Bench Hard," the misinformation rate, calculated by verifying individual claims, fell from the old model's 23.4% to 8.3%. The company describes this as a reduction of about 65%, but it is the result of a stress test designed to provoke wrong answers and does not guarantee that errors will fall by the same proportion in everyday use. Separately, in Artificial Analysis's "AA-Omniscience" test, the misinformation rate also improved from 53% to 44%. Although both tests show improvement, their evaluation conditions differ, so the figures cannot be compared directly.
API prices cut 33%, yet cost per task is roughly unchanged
Alongside the performance gains, Kimi K3.1 Pro also came with lower API prices. According to Artificial Analysis, the standard API's per-token prices for input, output, and cached input are all 33% lower than the old K3's. Yet the cost per task when the institute ran its evaluation tasks was $1.98, roughly unchanged from the previous model. The result shows that a cut in per-token price does not translate directly into lower actual work costs.
The table below summarizes pricing and usage conditions from Artificial Analysis's evaluation results and the official Kimi Code documentation. Note that pay-as-you-go API billing and the usage allowances in Kimi Code membership plans are separate pricing systems.
| Cost/speed item | Pricing and conditions |
|---|---|
| Standard API price per 1M tokens | $2 input, $10 output, $0.20 cached input |
| Artificial Analysis evaluation tasks | $1.98 per task. Output tokens rose from K3's 48K to 55K (a 14% increase, per the institute) |
| Kimi Code Highspeed | Speeds up model output. Available on Pro/Allegretto and above; consumes about twice the normal usage allowance |
A 33% drop in token price does not mean the cost of a single job falls by 33%. In Artificial Analysis's measurements, output tokens rose from about 48,000 for the old K3 to about 55,000. However, because a detailed breakdown including input tokens and cache usage is not available, it cannot be concluded that this increase in output alone kept costs flat. Also, the $1.98 figure is a measurement from the institute's evaluation tasks and is not an average cost common to all users and uses.
Moonshot describes Highspeed as two to three times faster than the standard version. According to the official Kimi Code documentation, however, only the model's response generation is sped up. Calling external tools and running scripts do not themselves get faster, so for work where those steps take time, overall turnaround will not necessarily shrink by the same factor. In addition, Highspeed consumes about twice the normal usage allowance, so the balance between speed and available processing volume must also be considered.
Using standard K3.1 Pro in Kimi Code requires a "Plus/Moderato" plan or above, while the 1-million-token context length and Highspeed require "Pro/Allegretto" or above. The model ID used in Code is k3.1-pro, which differs from the general API's kimi-k3.1-pro. In Kimi Code, thinking cannot be disabled, and the default reasoning effort is "max." When deciding whether to adopt or switch models, evaluation should consider not just unit price and generation speed but also the time spent on reruns, corrections, and checking deliverables in actual work.
Weights due by November 4; cyber capabilities deliberately restricted
Moonshot says it took measures to suppress cyberattack-related capabilities in developing this model. Specifically, it removed cyber-related content from the data used in mid-training and reinforcement learning while maintaining training that improves general programming ability. The decision was made to limit the risk that releasing the model would make advanced attack capabilities widely available. In the company's internal evaluation "xdev-bench-hard," pass@3, the success rate across three attempts, was 90.5% for the checkpoint before cyber-related data was removed, versus 81.0% for the final model.
This test, however, targets already-public vulnerabilities, giving the model things like source code, short hints, and executable binaries and verifying attack success in an isolated environment. The evaluation used a subset of randomly selected attempts and set no strict limits on execution time or output, so it cannot be directly compared with other benchmarks or real-world attack success rates. Moonshot itself also acknowledges that the final model's cyber-related capabilities are higher than the old K3's. The company says it aims to avoid a release that further raises frontier attack capabilities while gradually spreading capabilities useful for defense and auditing, but these measures cannot be regarded as a guarantee of safety.
At announcement, Kimi K3.1 Pro is available through Kimi.ai and the mobile app, as well as Work, Code, and the API. In enterprise admin settings, however, enabling by an administrator is off by default. In Kimi Code, starting a new session is recommended when switching models from an existing conversation. Because behavior may change due to protocol conversion at connection time, compatibility should be checked before introducing it into an existing development environment.
The key upcoming milestone is the release of the trained weights, planned by November 4, 2026. The assessment of "top open-weight model" that Artificial Analysis mentions presupposes that the weights are actually released and that the ranking holds at that point. After release, the focus will be on whether third parties can check the license and usage terms and reproduce performance in their own environments. Of particular importance is whether, in work involving long tool-driven operation, deliverable quality and total cost live up to the cloud-based evaluation results. Only once these can be verified will it be possible to judge concretely whether Kimi K3.1 Pro is worth moving into a different operating environment.