# Kimi K3.1 Pro is now on Tchavi

> Kimi K3.1 Pro is on Tchavi: 1M context, images and tools, at half the price of Opus 5.5. Verified benchmarks and limits measured on day one.

- Published: 2026-10-23
- Author: Gael — Tchavi founder
- Canonical: https://tchavi.com/en/blog/kimi-k3-1-pro-now-on-tchavi

---


Kimi K3.1 Pro, Moonshot AI's new model, is on Tchavi the day it was released, as `kimi-k3.1-pro`. On the benchmarks Moonshot published, it runs close to Claude Opus 5.5 on agentic coding, at about half its price on Tchavi. Before you wire it in, read the part about reasoning: one setting decides whether the same answer costs you 9 credits or 60.

## In short

- Tchavi, an OpenAI-compatible AI API gateway you pay for with mobile money, adds Kimi K3.1 Pro: `kimi-k3.1-pro`.
- 1,048,576-token context, image input, tool calling, strict JSON schemas.
- 4.39 credits per 1,000 input tokens and 21.95 per 1,000 output tokens: about half the price of Claude Opus 5.5.
- It reasons at maximum effort by default. `reasoning_effort: "none"` switches reasoning off; the levels in between are refused for now.

## What Moonshot reports

Moonshot presents K3.1 Pro as a post-trained successor to Kimi K3, built for long-horizon coding, office documents and finance. A few figures from its own chart, published with the model:

| Benchmark                         | Kimi K3.1 Pro | Claude Opus 5.5 | GPT-6 Astra |
| --------------------------------- | ------------- | --------------- | ----------- |
| Terminal-Bench 4.0                | 65.7          | 66.4            | 58.2        |
| SWE Marathon                      | 51.7          | 46.3            | 42.5        |
| Humanity's Last Exam (with tools) | 70.0          | 67.7            | 57.2        |
| OfficeQA Pro v2                   | 64.4          | 62.2            | 54.4        |
| OSWorld V2.1                      | 79.6          | 81.8            | 75.8        |
| Terminal-Bench Science 0.1        | 38.1          | 63.3            | 68.1        |

Two caveats. These are Moonshot's numbers, mostly measured in the Claude Code harness at maximum effort, and nobody has reproduced them yet. The chart also shows a clear weak spot: on terminal science work, K3.1 Pro trails far behind. Moonshot says it will release the weights by 4 November; for now, only Moonshot serves the model.

## The price on Tchavi

| Model                                              | Input (credits / 1K tokens) | Output (credits / 1K tokens) |
| -------------------------------------------------- | --------------------------- | ---------------------------- |
| [Kimi K3.1 Pro](https://tchavi.com/en/m/moonshot-ai/kimi-k3.1-pro)   | 4.39                        | 21.95                        |
| [Kimi K3](https://tchavi.com/en/m/moonshot-ai/kimi-k3)               | 6.6                         | 32                           |
| [Claude Opus 5.5](https://tchavi.com/en/m/anthropic/claude-opus-5-5) | 8.8                         | 44                           |
| [GPT-6 Astra](https://tchavi.com/en/m/openai/gpt-6-astra)            | 22                          | 110                          |

At October 2026 rates, KES 1,000 buys about 13,510 credits. A million tokens then costs roughly KES 325 of input and KES 1,625 of output. K3.1 Pro also costs less than Kimi K3, which stays in the catalogue.

## What I measured on the day we added it

In October 2026, I sent about thirty short requests through Tchavi before opening the model. Three results matter.

**Default reasoning is long.** With no setting, K3.1 Pro thinks at maximum effort. Asked to work out the margin of a cashew cooperative in Bohicon, Benin, it produced 2,411 reasoning tokens before a correct answer, in 38 seconds: 60 credits, about KES 4.50. That run is the demo on the [model page](https://tchavi.com/en/m/moonshot-ai/kimi-k3.1-pro). The same question in French hit the 3,000-token limit without answering. Set `max_tokens` to 8,000 or more.

**`"none"` switches reasoning off, and it works.** Moonshot's documentation says reasoning cannot be turned off, but with `reasoning_effort: "none"` the French question came back correct in 6 seconds, with no reasoning: 8.8 credits, under KES 1.

**The levels in between are broken.** With `minimal`, `low`, `medium` or `high`, the model wrote incoherent reasoning until it ran out of tokens, gave no answer, and each call was billed in full. So Tchavi refuses those values for free, with a 400 error that says what to do. I will reopen them once a new test shows the problem is fixed.

Everything else works as expected: images sent as base64, automatic or forced tool calls, `json_schema` and `json_object`. `temperature`, `top_p` and the penalties are ignored, because Moonshot fixes them. `logprobs` and `parallel_tool_calls` return an error.

## One call to try it

```bash
curl https://tchavi.com/api/v1/chat/completions \
  -H "Authorization: Bearer $TCHAVI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "kimi-k3.1-pro",
    "reasoning_effort": "none",
    "max_tokens": 1200,
    "messages": [
      {"role": "user", "content": "Work out the net margin: 1,200 kg of cashew nuts sold at KES 145 per kg, KES 128,000 of costs and 3% fees on the money collected."}
    ]
  }'
```

Remove the `reasoning_effort` line to let it reason at maximum effort, and raise `max_tokens` when you do. The cost of each call is in the `X-Credits-Used` response header. To get started, follow the [quickstart](https://tchavi.com/en/docs/quickstart) or the [chat completions](https://tchavi.com/en/docs/chat-completions) page.

## FAQ

### Does Kimi K3.1 Pro replace Kimi K3?

No, both stay in the catalogue. K3.1 Pro costs less. Keep K3 if you rely on `top_p` or the penalties, which K3.1 Pro ignores.

### Why does `reasoning_effort: "low"` return an error?

Because that value produced unusable output, billed in full. The error is free and tells you what to do: remove the field or set it to `"none"`.

### Does it read PDFs?

No. It reads text and images. For PDFs, use [Claude Sonnet 5](https://tchavi.com/en/m/anthropic/claude-sonnet-5) or [GPT-6 Sol](https://tchavi.com/en/m/openai/gpt-6-sol).

### Are its weights open?

Not yet. Moonshot says it will release them by 4 November 2026 and has not named the licence.

### Can I use it with the OpenAI SDK?

Yes. Set the base URL to `https://tchavi.com/api/v1`, use your Tchavi key and set `kimi-k3.1-pro` as the model.

[create your account and top up KES 250](https://tchavi.com/en/register?next=%2Fdashboard%2Fbilling%3Famount%3D250%26currency%3DKES%26from%3Dblog-kimi-k31-en)

> _This article is part of the series "Building an AI product from Africa". Tchavi is an OpenAI-compatible AI API gateway: one key gives access to over 100 text, image, audio and video models, and you pay for it with mobile money, no bank card ([countries and operators](https://tchavi.com/en/docs/payments)). More guides on [the blog](https://tchavi.com/en/blog). Follow us: [LinkedIn](https://www.linkedin.com/company/tchavi) and [X](https://x.com/tchaviai)._