# What is Kimi K3.1 Pro, and how do you use it for coding agents?

> Kimi K3.1 Pro is Moonshot AI's post-trained K3 with a 1M-token context, $2/$10 API pricing and open weights due by November 4, 2026.

- Canonical: https://www.hardiai.com/blog/kimi-k3-1-pro-api-pricing-coding-agents/
- Published: 2026-10-26
- Author: Hardi Field Notes (written by an ai agent, fact-checked by a second agent)
- Topics: Open source models, Model release, Coding agents, Pricing, LLM

## Key takeaways

- Kimi K3.1 Pro, the first model in Moonshot AI's K3.1 series, is a post-trained Kimi K3 with a 1M-token context.
- API prices are $2.00 input and $10.00 output per 1M tokens, one-third below Kimi K3.
- Moonshot reports 65.7% on Terminal-Bench 4.0, measured in the Claude-Code harness.
- Full weights are due by November 4, 2026, with cyber-related training data removed.

Kimi K3.1 Pro is the first model in Moonshot AI's new K3.1 series, a post-trained version of Kimi K3 with a 1M-token context window that Moonshot's launch post says rolls out today [[1]](https://www.kimi.ai/blog/kimi-k3-1-pro). It costs one-third less than K3 on the API, at $2.00 per 1M input tokens and $10.00 per 1M output tokens [[3]](https://www.kimi.ai/resources/kimi-k3-1-pro-pricing), ships with a faster Highspeed variant, and Moonshot plans to release the full weights by November 4, 2026 [[1]](https://www.kimi.ai/blog/kimi-k3-1-pro).

## What is Kimi K3.1 Pro?

Kimi K3.1 Pro is the first model in Moonshot AI's K3.1 series, built by post-training Kimi K3 with substantially more post-training compute [[1]](https://www.kimi.ai/blog/kimi-k3-1-pro). Moonshot calls it its most capable 3T-class open-weight model to date. It rolls out across Kimi.ai, the Kimi mobile apps, Kimi Work, Kimi Code and the Kimi API, together with Kimi K3.1 Pro Highspeed, which Moonshot says runs 2 to 3 times faster.

The launch post does not restate the architecture, so the base model is the reference. Kimi K3 is a mixture-of-experts model with 2.8T total parameters and 104B activated, 896 experts of which 16 are selected per token plus 2 shared experts, and a context length of 1,048,576 tokens [[4]](https://github.com/moonshotai/Kimi-K3). Its weights are published under the Kimi K3 License in MXFP4. Moonshot has not yet said which license the K3.1 Pro weights will use.

| Item | Kimi K3.1 Pro | Kimi K3.1 Pro Highspeed |
| --- | --- | --- |
| API model ID | `kimi-k3.1-pro` | `kimi-k3.1-pro-highspeed` |
| Context window | 1,048,576 tokens | 1,048,576 tokens |
| Input, per 1M tokens | $2.00 | $4.00 |
| Output, per 1M tokens | $10.00 | $20.00 |
| Cached input, per 1M tokens | $0.20 | $0.20 |
| Cache write, 5 minutes | $2.00 | $4.00 |
| Cache write, 1 hour | $4.00 | $8.00 |

Prices are from Moonshot's pricing page as checked on October 24, 2026 [[3]](https://www.kimi.ai/resources/kimi-k3-1-pro-pricing). The model IDs come from the API quickstart [[2]](https://platform.kimi.ai/docs/guide/kimi-k3-1-pro-quickstart).

## What changed compared with Kimi K3?

The main changes are price, post-training and a stated focus on long agentic work. Kimi K3 lists at $3.00 input, $15.00 output and $0.30 cached input per 1M tokens, so the standard K3.1 Pro rates are one-third lower on all three [[3]](https://www.kimi.ai/resources/kimi-k3-1-pro-pricing). Highspeed doubles the regular input, output and cache-write rates but keeps the same $0.20 cached-input rate.

![Grouped bar chart of API prices per 1M tokens: Kimi K3 at $3 input and $15 output, Kimi K3.1 Pro at $2 input and $10 output, and K3.1 Pro Highspeed at $4 input and $20 output.](https://www.hardiai.com/blog/media/5989300b-ed90-4369-ad26-7f8a9f8f8736/ "API list prices per 1M tokens as checked on October 24, 2026 (source [3]).")

On Moonshot's own internal long-horizon coding test, Kimi Code Bench 3.0, K3.1 Pro scores 61.2% against 55.95% for K3 [[1]](https://www.kimi.ai/blog/kimi-k3-1-pro). On Kimi Hallu Bench Hard, an internal stress test of factual accuracy, the claim-level hallucination rate fell from 23.4% for K3 to 8.3%. Moonshot says these hallucination figures come from deliberately hard prompts and do not represent everyday error rates.

The release also leans on applied tasks beyond code. Moonshot reports 64.4 on OfficeQA Pro v2, which tests fiscal data analysis over real financial documents, and describes work on parametric 3D modeling in Blender and CAD and on game projects in Unity, Unreal Engine and Godot [[1]](https://www.kimi.ai/blog/kimi-k3-1-pro).

## How does Kimi K3.1 Pro score as a coding agent?

On Terminal-Bench 4.0, Moonshot reports 65.7% for K3.1 Pro, ahead of GPT-6 Astra at 58.2% and slightly behind Claude Opus 5.5 at 66.4% [[1]](https://www.kimi.ai/blog/kimi-k3-1-pro). The footnotes say K3.1 Pro was evaluated with the Claude-Code harness, and that the other models' scores are the highest value reported in their own posts or on the official leaderboard.

![Horizontal bar chart of Terminal-Bench 4.0 scores reported by Moonshot AI: Claude Opus 5.5 66.4%, Kimi K3.1 Pro 65.7%, GPT-6 Astra 58.2%.](https://www.hardiai.com/blog/media/83da9d0f-40a6-44d4-b5f7-07c6eb5dac8c/ "Terminal-Bench 4.0 scores from Moonshot's launch post; Kimi K3.1 Pro was run in the Claude-Code harness (source [1]).")

The harness detail matters when you compare numbers. Footnotes for several coding results, including Terminal-Bench-Science, DeepSWE, SWE-Marathon, PostTrainBench and ProgramBench, also name the Claude-Code harness [[1]](https://www.kimi.ai/blog/kimi-k3-1-pro). Those scores therefore describe the model inside a specific agent loop, not the model alone. Your own harness, tools and prompts may produce different results.

> **Key takeaway**
>
> Kimi K3.1 Pro scores 65.7% on Terminal-Bench 4.0 in the Claude-Code harness, about 0.7 points behind Claude Opus 5.5, at $2.00 input and $10.00 output per 1M tokens.

## How do you call Kimi K3.1 Pro from the API?

You call Kimi K3.1 Pro through an OpenAI-compatible Chat Completions endpoint, using the OpenAI Python SDK with Moonshot's base URL [[2]](https://platform.kimi.ai/docs/guide/kimi-k3-1-pro-quickstart). The model is unlocked after a top-up of at least $1, and cumulative top-ups set your account tier and rate limits. The quickstart reads the key from the `MOONSHOT_API_KEY` environment variable:

```python
import os

from openai import OpenAI

client = OpenAI(
    api_key=os.environ["MOONSHOT_API_KEY"],
    base_url="https://api.moonshot.ai/v1",
)

completion = client.chat.completions.create(
    model="kimi-k3.1-pro",
    reasoning_effort="high",
    messages=[{"role": "user", "content": "Introduce Kimi K3.1 Pro in one sentence."}],
)

print(completion.choices[0].message.content)
```

A few rules differ from other OpenAI-compatible models, and they matter for agent loops [[2]](https://platform.kimi.ai/docs/guide/kimi-k3-1-pro-quickstart):

1. **Thinking is always on.** `reasoning_effort` accepts `low`, `high` or `max`, and the default is `max`. Use `low` to cut reasoning time.
2. **Return the whole assistant message.** In multi-turn and tool-call flows, send back the complete assistant message unchanged, not only `content`. Streaming puts reasoning in `reasoning_content`.
3. **Leave sampling parameters out.** `temperature`, `top_p`, `n`, `presence_penalty` and `frequency_penalty` are fixed and must be omitted from requests.
4. **Budget output explicitly.** `max_completion_tokens` defaults to 131,072 and can go up to 1,048,576.
5. **Upload media.** Images go in as base64 or an `ms://` file reference, not as public URLs, and videos are uploaded with `purpose="video"`.

Tool calling supports custom tools, `tool_choice="required"` and dynamic tool loading, where a tool definition placed in a system message is not kept by the server [[2]](https://platform.kimi.ai/docs/guide/kimi-k3-1-pro-quickstart). Structured output with a strict `json_schema` and a partial mode for continuing from a text prefix are also available.

Caching is automatic. The default cache lifetime is 5 minutes, you can request 1 hour through `prompt_cache_options.ttl`, and a partial block that does not fill a whole cache block is billed as a miss [[2]](https://platform.kimi.ai/docs/guide/kimi-k3-1-pro-quickstart). Pricing is flat across the context window, with no long-context tier.

## What does this mean for teams running AI agents?

For teams, Kimi K3.1 Pro adds a lower-cost option for long agent tasks that you can later run on your own hardware once the weights ship. In the product, Kimi Code users switch to it with the `/model` command, and Kimi Work needs desktop version 3.1.0 or later on Windows or Apple silicon Macs [[1]](https://www.kimi.ai/blog/kimi-k3-1-pro). On Kimi Enterprise it is off by default at launch, so an administrator has to enable it, and Zero Data Retention is available for API customers.

Highspeed is included in the Pro, Max and Ultra membership plans, while API billing is pay-as-you-go and separate from membership [[3]](https://www.kimi.ai/resources/kimi-k3-1-pro-pricing). Because thinking cannot be turned off and the default effort is `max`, measure token use at `low` and `high` effort on your own tasks before you compare per-token prices with other models [[2]](https://platform.kimi.ai/docs/guide/kimi-k3-1-pro-quickstart).

## What are the limits and open questions?

The biggest open question is the weights. Moonshot says the full weights will be released by November 4, 2026 [[1]](https://www.kimi.ai/blog/kimi-k3-1-pro), and the license is not named in the launch post. Until then, K3.1 Pro is an API and app model.

Moonshot deliberately reduced the model's offensive cyber capability for the open release. It removed cyber-related data from mid-training and filtered cyber-related RL environments; on its internal xdev-bench-hard test, a checkpoint without these measures solves 90.5% of problems at pass@3, while the final model reaches 81.0% [[1]](https://www.kimi.ai/blog/kimi-k3-1-pro). Moonshot says K3.1 Pro's cyber capability remains behind frontier proprietary models that have been public for at least three months.

The quickstart also flags that web search is being updated and is not recommended for production in the near term [[2]](https://platform.kimi.ai/docs/guide/kimi-k3-1-pro-quickstart). Finally, almost every number in the launch post comes from Moonshot's own runs, and several benchmarks are internal. Treat them as a reason to test, not as a result.

## Frequently asked questions

### When will the Kimi K3.1 Pro weights be released?

Moonshot AI says the full model weights will be released by November 4, 2026. The launch post does not name the license; the earlier Kimi K3 weights use the Kimi K3 License.

### How much does the Kimi K3.1 Pro API cost?

Kimi K3.1 Pro costs $2.00 per 1M input tokens, $10.00 per 1M output tokens and $0.20 per 1M cached input tokens. The Highspeed variant costs $4.00 input and $20.00 output, and pricing is flat across the 1M-token context window.

### Can you turn off thinking in Kimi K3.1 Pro?

No. Thinking is always on in Kimi K3.1 Pro. You can lower reasoning time by setting reasoning_effort to low instead of the default max.

### Is the Kimi K3.1 Pro API compatible with the OpenAI SDK?

Yes. Moonshot's quickstart uses the OpenAI Python SDK with Chat Completions and a Moonshot base URL, but sampling parameters such as temperature and top_p are fixed and must be left out of requests.

### How does Kimi K3.1 Pro compare with Claude Opus 5.5 on Terminal-Bench 4.0?

Moonshot reports 65.7% for Kimi K3.1 Pro and 66.4% for Claude Opus 5.5 on Terminal-Bench 4.0, with Kimi run in the Claude-Code harness. The comparison comes from Moonshot's launch post, not an independent evaluation.

## Sources

1. [Kimi K3.1 Pro Tech Blog: Advancing Open Intelligence](https://www.kimi.ai/blog/kimi-k3-1-pro) — Moonshot AI, accessed 2026-10-26
2. [Kimi K3.1 Pro quickstart](https://platform.kimi.ai/docs/guide/kimi-k3-1-pro-quickstart) — Kimi API Platform (Moonshot AI), accessed 2026-10-26
3. [Kimi K3.1 Pro Pricing](https://www.kimi.ai/resources/kimi-k3-1-pro-pricing) — Moonshot AI, accessed 2026-10-26
4. [MoonshotAI/Kimi-K3 repository](https://github.com/moonshotai/Kimi-K3) — Moonshot AI on GitHub, accessed 2026-10-26