# Kimi K3.1 Pro: Near-Frontier Open Weights at $2 per Million Tokens

> Moonshot's Kimi K3.1 Pro is a 3T-class open-weight model at $2 per million input tokens. What the price, benchmarks and caveats mean for indie founders.

- Canonical: https://zeroraised.com/blog/kimi-k3-1-pro-open-weights
- Published: 2026-10-26
- Author: ZeroRaised Team (https://zeroraised.com/blog/authors/zeroraised)
- Publisher: ZeroRaised (https://zeroraised.com/)
- Section: Growth
- Tags: artificial-intelligence, developer-tools, bootstrapping

Kimi K3.1 Pro is Moonshot AI's most capable open-weight model to date, and the price should interest a founder: $2.00 per million input tokens and $10.00 per million output tokens, with cached input at $0.20 ([Kimi pricing](https://www.kimi.ai/resources/kimi-k3-1-pro-pricing)). Moonshot reports that it lands within a point of Claude Opus 5.5 on an agentic coding benchmark while costing half as much per token ([Simon Willison](https://simonwillison.net/2026/oct/23/kimi-k31-pro/)). That is real pressure on the closed labs, even if the benchmarks are the vendor's own and the weights are still a promise.

**TL;DR**

- K3.1 Pro is a 3T-class open-weight model with a 1,048,576-token context and flat per-token pricing ([Kimi pricing](https://www.kimi.ai/resources/kimi-k3-1-pro-pricing)).
- Moonshot's Terminal-Bench 4.0 result is 65.7%, between GPT-6 Astra at 58.2% and Claude Opus 5.5 at 66.4% ([Kimi blog](https://www.kimi.ai/blog/kimi-k3-1-pro)).
- Full weights are promised by 4 November 2026; until then it is a hosted API ([Kimi blog](https://www.kimi.ai/blog/kimi-k3-1-pro)).
- Artificial Analysis scored it 55, behind only Opus 5.5 at 58, per Simon Willison ([Simon Willison](https://simonwillison.net/2026/oct/23/kimi-k31-pro/)).

**Disclosure:** ZeroRaised publishes this analysis and is a launch site for bootstrapped products. Facts come from the linked sources as checked on 26 Oct 2026; ZeroRaised has no commercial relationship with the companies named.

**An open-weight model** is a model whose trained parameters are published for others to download, run and fine-tune, as distinct from a closed model that you can only call through an API. Open weight is not the same as open source: the licence and the release of training code or data decide that.

## What did Moonshot actually release?

Moonshot describes K3.1 Pro as "our most capable 3T-class open-weight model to date, with a 1-million-token context window" ([Kimi blog](https://www.kimi.ai/blog/kimi-k3-1-pro)). The pricing page lists the context at 1,048,576 tokens ([Kimi pricing](https://www.kimi.ai/resources/kimi-k3-1-pro-pricing)). A Highspeed variant is said to deliver two to three times the speed ([Kimi blog](https://www.kimi.ai/blog/kimi-k3-1-pro)).

The catch is timing. Moonshot says full model weights will be released by 4 November 2026 ([Kimi blog](https://www.kimi.ai/blog/kimi-k3-1-pro)). Until then K3.1 Pro is a hosted API: you can call it, but not download it, inspect it or run it yourself. Treat self-hosting as a future option, not a current one. Moonshot also says it removed cyber-related data from training for the open-weight release ([Kimi blog](https://www.kimi.ai/blog/kimi-k3-1-pro)), so the weights you eventually get may differ from the hosted model.

## The price is the story

At $2.00 input and $10.00 output per million tokens, K3.1 Pro is one third cheaper than Kimi K3, listed at $3.00 and $15.00 ([Kimi pricing](https://www.kimi.ai/resources/kimi-k3-1-pro-pricing)). Cached input is $0.20, and the Highspeed variant is $4.00 and $20.00 with the same cached rate ([Kimi pricing](https://www.kimi.ai/resources/kimi-k3-1-pro-pricing)).

Simon Willison notes that $2/$10 is "the same as GPT-6.1 Sol and half the price of Claude Opus 5.5" ([Simon Willison](https://simonwillison.net/2026/oct/23/kimi-k31-pro/)). Moonshot also keeps pricing flat, with no context-length tiers ([Kimi pricing](https://www.kimi.ai/resources/kimi-k3-1-pro-pricing)).

One billing detail matters more than the headline rate: K3.1 Pro always reasons, and the reasoning effort setting supports low, high and max, with max as the default ([Kimi pricing](https://www.kimi.ai/resources/kimi-k3-1-pro-pricing)). Reasoning tokens are output tokens, so leaving the default on every request means paying for thinking that some tasks did not need. Our guide to [reducing LLM costs](https://zeroraised.com/blog/how-to-reduce-llm-costs) covers effort settings, caching and batching.

## How good is it, really?

On Terminal-Bench 4.0, Moonshot reports 65.7% for K3.1 Pro, against 58.2% for GPT-6 Astra and 66.4% for Claude Opus 5.5 ([Kimi blog](https://www.kimi.ai/blog/kimi-k3-1-pro)). Those are vendor-reported figures, so read them as Moonshot's claim rather than an independent result.

Moonshot also reports an internal hallucination stress test where the claim-level rate fell from 23.4% for K3 to 8.3% for K3.1 Pro ([Kimi blog](https://www.kimi.ai/blog/kimi-k3-1-pro)). The company says these are stress tests, not everyday error rates, so do not quote 8.3% as a general reliability number.

The stronger signal comes from outside. Artificial Analysis scored Kimi at 55, behind only Claude Opus 5.5 at 58 and ahead of Claude Fable 5.1, GPT-6 Astra and Gemini 4 Argon, all at 53 ([Simon Willison](https://simonwillison.net/2026/oct/23/kimi-k31-pro/)). Willison calls it "the first time a (soon-to-be) open weight model has scored this highly against the frontier", and notes the weights had not been released as of 23 October ([Simon Willison](https://simonwillison.net/2026/oct/23/kimi-k31-pro/)).

## The caveats that should slow you down

**The benchmarks are vendor-reported.** Terminal-Bench 4.0 is Moonshot's own table ([Kimi blog](https://www.kimi.ai/blog/kimi-k3-1-pro)). The independent score is close, which helps, but your workload is the only benchmark that matters.

**The weights are not out yet.** As of 23 October they had not been released, and the promised date is 4 November 2026 ([Simon Willison](https://simonwillison.net/2026/oct/23/kimi-k31-pro/); [Kimi blog](https://www.kimi.ai/blog/kimi-k3-1-pro)). A date is not a download.

**The launch has a reputational problem.** VentureBeat reported allegations that influencers were asked to promote the model without disclosure ([VentureBeat](https://venturebeat.com/technology/moonshots-kimi-k3-1-pro-flirts-with-open-weights-leadership-amid-paid-influencer-scandal-allegations)). That does not change the API, but it is a reason to test the model yourself.

**Procurement risk is real for Chinese models.** The US order excluding DeepSeek from federal civilian systems also covers third-party products built on its technology ([CNBC](https://www.cnbc.com/2026/10/20/us-deepseek-federal-agency-ban.html)). It names DeepSeek, not Moonshot, but shows how fast a model can become a compliance problem for anyone selling to government buyers.

Moonshot says Zero Data Retention is available for API customers ([Kimi blog](https://www.kimi.ai/blog/kimi-k3-1-pro)), but availability is not the same as a default. Confirm the terms before customer data goes through it.

## Should you switch?

Not for the price alone. Switch if a coding or agent workload passes your own evaluation on K3.1 Pro and the saving is material after reasoning tokens. It is strongest in agentic coding, long context and tool use.

Keep the integration thin. Put model calls behind one internal interface, keep prompts in your repository and hold a few real tasks with known good answers. That is the same groundwork that makes any provider change easier, and it is covered in [how to launch an AI app](https://zeroraised.com/blog/how-to-launch-an-ai-app). The [OpenAI and Anthropic discount war](https://zeroraised.com/blog/openai-anthropic-discount-war) is the other half of this price story.

## What this means for bootstrapped founders

- **Price for output, not input.** Reasoning is on by default at max effort, so the output bill is the one that can surprise you ([Kimi pricing](https://www.kimi.ai/resources/kimi-k3-1-pro-pricing)). Set effort per feature.
- **Run a five task eval before you commit.** Take five real requests, write down the correct output and score both models. That beats any leaderboard.
- **Design for two providers.** A single abstraction layer turns a price change or a compliance problem into a config edit.
- **Do not promise self-hosting yet.** Wait for the 4 November release and check the licence first.
- **Use the price cut to protect margin.** If unit economics were tight on K3, the reduction is room to fund support rather than a reason to cut your price. Our [SaaS pricing guide](https://zeroraised.com/blog/how-to-price-a-saas-product) covers the trade-off.

## Where ZeroRaised fits

ZeroRaised is a launch site for products built without institutional VC or accelerator equity. If you have shipped something that uses a model like this, you can [submit your product](https://zeroraised.com/submit) and get a public page that states what it does and who made it. There is no promise of traffic. To follow the next analysis, [join the ZeroRaised newsletter](https://zeroraised.com/newsletter).

## Sources

- [Kimi K3.1 Pro announcement](https://www.kimi.ai/blog/kimi-k3-1-pro)
- [Kimi K3.1 Pro pricing](https://www.kimi.ai/resources/kimi-k3-1-pro-pricing)
- [VentureBeat on Kimi K3.1 Pro](https://venturebeat.com/technology/moonshots-kimi-k3-1-pro-flirts-with-open-weights-leadership-amid-paid-influencer-scandal-allegations)
- [Simon Willison on Kimi K3.1 Pro](https://simonwillison.net/2026/oct/23/kimi-k31-pro/)
- [CNBC on the DeepSeek federal agency ban](https://www.cnbc.com/2026/10/20/us-deepseek-federal-agency-ban.html)

## Frequently asked questions

### Is Kimi K3.1 Pro actually open source?

Not yet, and open weight is not the same as open source. Moonshot says the full model weights will be released by 4 November 2026, but until then K3.1 Pro is available as a hosted API. Open weights let you download and run the parameters; open source usually also means the training code and data are released.

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

The standard API is $2.00 per million input tokens and $10.00 per million output tokens, with cached input at $0.20. The Highspeed variant is $4.00 input and $20.00 output. Moonshot lists a flat per-token price with no context-length tiers, checked on 24 October 2026.

### Is Kimi K3.1 Pro better than Claude Opus 5.5?

On Moonshot's own Terminal-Bench 4.0 figures it is slightly behind, 65.7% against 66.4% for Claude Opus 5.5. Those numbers are vendor-reported. A third-party index from Artificial Analysis put Kimi at 55 against 58 for Opus 5.5. Test both on your own tasks before switching.

### Should a bootstrapped founder build on Kimi K3.1 Pro?

It is worth testing if cost and vendor independence matter to you, especially for coding and agent workflows. Keep the integration thin enough to swap providers, check data retention terms, and do not promise customers self-hosting until the weights are actually published.