Kimi K3.1 Pro is Moonshot AI's latest flagship reasoning model and the first release in the K3.1 series. Released on October 23, 2026, it upgrades Kimi K3 through substantially more post-training compute, with improvements focused on long-horizon coding, AI agents, 3D creation, and financial research.
It keeps K3's roughly one-million-token context window and native visual understanding. Moonshot offers a standard version and a Highspeed tier that delivers approximately two to three times the output speed. Developers can now test the Kimi K3.1 Pro API through GPTProto.
Status on October 26, 2026: The hosted model and API are live. Full weights are scheduled for release by November 4, 2026.
Kimi K3.1 Pro at a Glance
| Specification |
Kimi K3.1 Pro |
| Developer |
Moonshot AI |
| Release date |
October 23, 2026 |
| Position |
First model in the K3.1 series |
| Foundation |
Post-trained upgrade to Kimi K3 |
| Model class |
3-trillion-parameter class |
| Context window |
1,048,576 tokens |
| Input / output |
Text and visual input / text output |
| Reasoning |
Always on; low, high, or max effort |
| Official API model ID |
kimi-k3.1-pro |
| Official API price |
$2 input / $10 output per 1M tokens |
| Open-weight status |
Full weights scheduled by November 4, 2026 |
| Highspeed option |
Same capabilities at approximately 2–3× output speed |
K3.1 Pro is not a larger-context version of K3 or a completely new foundation. Its main upgrade is the additional post-training and reinforcement learning applied to complex tasks.
When Was Kimi K3.1 Pro Released?
The official Kimi Code release log dates the rollout to October 23, 2026. Some international reports show October 24 because of publication timing and time zones, but October 23 is the clearest release date to use.
K3.1 Pro is available on Kimi.com, Kimi's apps, Kimi Work, Kimi Code, and the Kimi API. Moonshot calls it an open-weight model, but its weights were not public on this article's October 26 cutoff date. Hosted access is live; self-hosting remains pending.
What's New in Kimi K3.1 Pro vs Kimi K3?
K3.1 Pro uses K3 as its foundation rather than replacing it with a larger context window. The upgrade centers on reported task performance, lower launch API prices, and more post-training for coding, agents, 3D work, and finance.
| Difference |
Kimi K3.1 Pro |
Kimi K3 |
| Position |
Current K3.1 flagship |
Previous-generation flagship |
| Context |
1,048,576 tokens |
1,048,576 tokens |
| Native vision |
Yes |
Yes |
| Kimi Code Bench 3.0 Long-horizon |
61.2% |
55.95% |
| Official input/output price |
$2 / $10 |
$3 / $15 |
| Weights |
Scheduled by November 4 |
Already released |
| Best fit |
New coding, research, and agent deployments |
Existing integrations and current self-hosting |
For teams already using the Kimi K3 API, the reason to upgrade is better reported performance and lower hosted pricing—not a bigger context window.
Four Key Kimi K3.1 Pro Features
Long-Horizon Coding and Agents
K3.1 Pro targets work in which an agent must inspect information, call tools, revise a plan, and continue for many steps. Moonshot demonstrates repository fixes, project builds, game development, and extended engineering experiments. The surrounding agent still needs to provide tools, permissions, tests, and recovery rules.
3D Creation and Visual-to-Code
This is not a conventional text-to-3D generator. K3.1 Pro interprets text, sketches, screenshots, or reference images, then creates code or scripts for Blender, CAD software, Three.js, and game engines. External software produces and renders the final asset.
Financial Research and Knowledge Work
The model is aimed at workflows combining document retrieval, cross-document checks, spreadsheet logic, calculations, and written conclusions. Potential uses include financial modeling, filing analysis, investment research, and business operations. Every material number and source still requires verification.
1M Context and Always-On Reasoning
The prompt, conversation, tool results, reasoning, and answer share the 1,048,576-token window. Reasoning cannot be disabled, although developers can choose low, high, or max effort. Multi-turn applications must also retain the complete assistant message and tool state.
Kimi K3.1 Pro Benchmarks: What Do They Show?
In Moonshot's launch evaluation, K3.1 Pro approached current proprietary frontier models on several coding and knowledge-work tests, but it did not lead every category.
| Benchmark |
K3.1 Pro |
Relevant comparison |
| Terminal-Bench 4.0 |
65.7% |
Opus 5.5: 66.4%; Astra: 58.2% |
| Kimi Code Bench 3.0 Long-horizon |
61.2% |
Kimi K3: 55.95% |
| BenchCAD with tools |
92.7% |
Opus 5.5: 96.2%; Astra: 95.9% |
| OfficeQA Pro v2 |
64.4 |
Led Moonshot's launch comparison |
| Terminal-Bench Science 0.1 |
38.1% |
Astra: 68.1%; Opus 5.5: 63.3% |
Moonshot ran K3.1 Pro at max reasoning effort, and several coding tests used the Claude Code harness. Competitor scores were not always produced with identical harnesses, safeguards, or reasoning settings, so this is not a perfectly controlled head-to-head test.
In Artificial Analysis' independent evaluation, K3.1 Pro scored 55 on the Intelligence Index, versus 58 for Claude Opus 5.5 Max. It also generated about 150 million output tokens across the evaluation, compared with a median of approximately 87 million. The result supports near-frontier capability, but also suggests unusually high reasoning and output use.
Kimi K3.1 Pro Pricing and Highspeed Explained
| Version |
Input / 1M |
Output / 1M |
Cache hit / 1M |
Recommended use |
| Kimi K3.1 Pro |
$2 |
$10 |
$0.20 |
Batch, background, and cost-aware work |
| Kimi K3.1 Pro Highspeed |
$4 |
$20 |
$0.20 |
Interactive coding and sequential agents |
| Kimi K3 |
$3 |
$15 |
$0.30 |
Existing K3 deployments |
Highspeed is not a quality upgrade. Moonshot describes it as the same model capabilities at approximately two to three times the output speed. This does not guarantee that every request will finish in one-third of the time because time to first token, tool execution, network delay, and reasoning length still affect total latency.
Use standard K3.1 Pro for unattended research and background jobs. Consider the Kimi K3.1 Pro Highspeed API when a user is waiting for each answer or every agent step blocks the next.
Kimi K3.1 Pro vs Opus 5.5, Astra, and GLM 5.3
Kimi K3.1 Pro vs Claude Opus 5.5
Standard K3.1 Pro costs half as much per input and output token as Opus 5.5 at official list rates. Opus currently has a slightly higher independent Intelligence Index score and a more mature native Claude Code ecosystem. Choose the Claude Opus 5.5 API for high-risk coding and final review; test Kimi first when cost, visual input, or the future open-weight route matters more.
Is Kimi K3.1 Pro Better Than GPT-6 Astra?
There is no universal winner. K3.1 Pro leads Astra in Moonshot's Terminal-Bench 4.0 result and costs much less per token. Astra leads Moonshot's scientific-agent and tool-assisted CAD results and is built for computer use. The GPT-6 Astra API is the stronger starting point when direct computer interaction is central.
Kimi K3.1 Pro vs GLM-5.3
Both provide approximately one million tokens of context and always-on reasoning. Kimi adds native visual understanding and a stronger emphasis on 3D work. GLM-5.3 is text-only but has lower official rates of $1.40 input and $4.40 output per million tokens. See What Is GLM-5.3? for its confirmed capabilities and pricing.
Can You Use Kimi K3.1 Pro with Claude Code and Codex?
Yes, but the model is connected through Kimi's service; it is not built into Anthropic's or OpenAI's model catalog.
Kimi publishes an official Claude Code setup guide and a separate Codex configuration guide. Kimi Code supports the OpenAI Responses API, allowing Codex clients to connect through a custom provider without a local protocol converter.
The Codex route currently documents text and image input rather than direct video input. Developers should also confirm the correct model ID for their access channel because the hosted API and Kimi Code plans do not use identical names in every configuration.
Early Feedback and Launch Limitations
In one same-prompt 3D agent experiment, K3.1 Pro completed a Three.js scene with fewer agent steps than Astra and created its own browser test script. However, an incompatible temperature setting broke its visual-review calls, showing how strongly agent configuration can affect the result.
Other users reported rapid subscription-quota consumption and long outputs. Reuters reported that Moonshot acknowledged launch-related rate limiting and unnecessary quota consumption. The company reset part of subscribers' allowances and adjusted parameters to reduce unnecessary subagent calls and output length.
Subscription quotas and pay-as-you-go API billing are separate. Developers should measure successful-task cost, reasoning tokens, retries, and elapsed time rather than judging the model only by its listed token rate.
Should You Use Kimi K3.1 Pro?
Choose K3.1 Pro for large codebases, visual-to-code work, financial documents, or long-running agents that can be checked with tests and acceptance criteria. Consider Opus 5.5 for high-risk coding, Astra for computer-use workflows, GLM-5.3 for cheaper text-only tasks, and Highspeed when output latency matters more than token cost.
Kimi K3.1 Pro is a meaningful K3 upgrade, not a completely new foundation. Its combination of improved long-horizon performance, native vision, a one-million-token context window, and lower hosted pricing makes it competitive with frontier models. The best choice still depends on the task, agent harness, token efficiency, and required latency.
Try the Kimi K3.1 Pro API on GPT Proto, or choose Kimi K3.1 Pro Highspeed when faster output matters more than token cost.
