Kimi K3.1 Pro: Price, Benchmarks and the Nov 4 Open-Weights Date

Kimi K3.1 Pro: Price, Benchmarks and the Nov 4 Open-Weights Date

Kimi K3.1 Pro is Moonshot AI's new flagship, released on October 23, 2026. It costs $2 per million input tokens and $10 per million output tokens on the Kimi API, a third less than Kimi K3, and Artificial Analysis scores it 55 on its Intelligence Index, three points behind Claude Opus 5.5. Moonshot says the full weights will be published by November 4, 2026. If that happens, it will be the strongest model anyone can download.

Until then it is an API model at exactly the list price of Claude Sonnet 5.5. This guide covers what you pay, which benchmark numbers come from Moonshot and which come from an independent lab, what we measured when we ran it ourselves, and what the November 4 release will and won't include. Every price and spec below was read on Moonshot's, Anthropic's and Artificial Analysis's own pages on October 25, 2026.

TL;DR

• Price: $2 / $10 per million tokens, cached input $0.20. Kimi K3 was $3 / $15. Highspeed is $4 / $20.
• Independent score: 55 on the Artificial Analysis Intelligence Index, 4th of 215 models listed. Opus 5.5 is 58; Sonnet 5.5 and Muse Wonder Full are 56.
• Cheaper per task than its list price suggests: Artificial Analysis spent $2.11 per index task on it against $5.46 for Sonnet 5.5, at the same token prices, because K3.1 Pro writes far fewer output tokens.
• Moonshot's own numbers: 65.7% on Terminal-Bench 4.0 against 66.4% for Opus 5.5. Its pages disagree with each other in places (details below), so treat them as vendor claims.
• Open weights: "by Nov 4, 2026". License not announced yet. The open release has cyber-related training data removed.

How much does Kimi K3.1 Pro cost?

Moonshot's pricing page lists three K3-series models. All of them have a 1,048,576-token context window.

Source: platform.kimi.ai/docs/pricing/chat, read October 25, 2026.

Three details worth knowing before you budget:

1. A 5-minute cache write costs the same as plain input. You pay nothing extra to write a prefix into the default cache, then $0.20 per million tokens on every hit. Long system prompts and tool lists get cheap fast.
2. K3.1 Pro is locked until you top up. Moonshot calls it a flagship model, unlocked after a successful top-up of at least $1. Your cumulative top-up also sets your rate-limit tier.
3. Highspeed keeps the cached-input price. It doubles input and output prices but cache hits stay at $0.20, so a cache-heavy agent pays less than double for the speed.

How does that compare with Claude?

Claude prices from Anthropic's pricing page, read October 25, 2026.

So the headline "half the price of Opus 5.5" is true, and so is the less exciting framing: K3.1 Pro costs the same per token as Sonnet 5.5, and Sonnet's cache reads are half the price. Where Kimi wins is how many tokens it spends. More on that below.

How good is Kimi K3.1 Pro? Independent vs vendor benchmarks

What Artificial Analysis measured

Artificial Analysis runs the same test suite on every model and pays for the tokens. On October 25, 2026 its leaderboard showed:

Two things stand out. K3.1 Pro gained 11 points over Kimi K3 for about the same cost per task. And at identical token prices it ran the index for 39% of what Sonnet 5.5 cost, because it used about a third as many output tokens. The trade is speed: 73.5 tokens per second against Sonnet's 132.7.

Simon Willison wrote on launch day that K3.1 Pro sat behind only Opus 5.5. The leaderboard we read two days later also lists Sonnet 5.5 and Muse Wonder Full at 56, so "one of the top five" is the accurate version today.

What Moonshot claims

Moonshot's launch post reports these numbers, all at max reasoning effort:

• Terminal-Bench 4.0: 65.7%, against 58.2% for GPT-6 Astra and 66.4% for Claude Opus 5.5 (run with the Claude Code harness).
• Cost per Terminal-Bench task, from the post's chart: $5.25 for K3.1 Pro, $14.30 for Opus 5.5, $22.22 for Sonnet 5.5.
• Hallucination rate on its internal Kimi Hallu Bench Hard: down from 23.4% for K3 to 8.3%.
• Internal finance evals: +7.2 points on financial modeling and +11.6 on financial research over K3.

We'd hold these loosely, for a reason you can check yourself. On October 25, Moonshot's pages didn't agree with each other:

• The post's text puts Opus 5.5 at 66.4% on Terminal-Bench 4.0. The chart in the same post plots Opus 5.5 at 64.85%.
• The post gives a hallucination rate of 8.3%. The API quickstart gives 6.85% for the same drop from K3.
• The quickstart still has unfilled template placeholders where the Terminal-Bench scores should be ("it scores [xx.x%]").

None of that means the model is weak; the independent score says it isn't. It means the vendor table was published in a hurry, and you should benchmark on your own tasks before you switch.

What we measured: does the reasoning effort setting matter?

K3.1 Pro always thinks. You can't turn reasoning off, only set reasoning_effort to low, high or max, and the default is max. We wanted to know what that default costs on the short, everyday calls most automations make.

We ran eight short tasks with automatic checks through OpenRouter on October 25, 2026: two arithmetic questions, a letter count, a trick date question (February 29, 2026 doesn't exist), an invoice-to-JSON extraction, and three Python functions that we executed against unit tests. Each task ran once on K3.1 Pro at every effort level (served by Moonshot itself) and once on Claude Sonnet 5.5 at its default. That's 32 calls and $0.13 in total, billed amounts taken from OpenRouter's usage.cost. The script and raw results are public in /benchmarks/kimi-k3-1-pro-effort/.

What we take from it:

• The default costs you 2.3x on easy calls for no accuracy gain. max used nearly three times the reasoning tokens of low and every answer was already right at low. For classification, extraction and other short calls, set reasoning_effort="low" explicitly.
• On short tasks Sonnet 5.5 was cheaper and 2-3x faster. Sonnet skipped thinking on 5 of 8 tasks; K3.1 Pro always thinks, even at low. K3.1 Pro also counted more input tokens for the same prompts (1,079 vs 628 across the eight), so some per-request overhead is added on Moonshot's side.
• That's the opposite of the long-task picture. On Artificial Analysis's much harder suite, K3.1 Pro cost less than half as much as Sonnet 5.5 per task. Short calls favour Sonnet, long reasoning-heavy tasks favour Kimi. Test on the kind of work you actually run.

Eight tasks is a sanity check, not a benchmark: one run each, easy enough that all four setups got everything right. It measures the cost of the effort setting, not intelligence.

What will the November 4 open weights include?

Moonshot's wording is "The full model weights will be released by Nov 4, 2026." Until they're out, Artificial Analysis lists K3.1 Pro as proprietary, and so does our model directory entry. What we know so far:

• Same size as Kimi K3. K3.1 Pro is a post-trained version of K3, a 2.8-trillion-parameter mixture-of-experts model with 104 billion parameters active per token. Moonshot calls it "our most capable 3T-class open-weight model to date."
• Cyber-related data removed. Moonshot says it removed cyber-related data from both mid-training and RL for the open-weight release, while keeping general coding tasks. On its internal vulnerability benchmark (xdev-bench-hard, pass@3), a checkpoint trained without those controls solved 90.5% of tasks; the final K3.1 Pro solves 81.0%. Moonshot's stated rule is not to "release open weights that would extend the absolute frontier of cyber capability."
• License: not announced. Kimi K3 shipped under a custom MIT-derived license that requires a separate agreement for model-as-a-service businesses above $20 million in revenue. We cover those terms in Kimi K3's open weights: read the license first. Don't assume K3.1 Pro gets the same terms until Moonshot publishes them.

Self-hosting a 2.8T-parameter model is a cluster project, not a workstation one. For most teams the practical effect of open weights is more hosting providers and lower prices on OpenRouter-style marketplaces, plus the option to run it inside your own cloud for data-residency reasons. Our guide to open-weight vs closed models walks through when that trade is worth it.

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

The Kimi API is OpenAI-compatible. Point the OpenAI SDK at Moonshot's base URL and use the model ID kimi-k3.1-pro (or kimi-k3.1-pro-highspeed):

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",  # low | high | max (default max)
    messages=[{"role": "user", "content": "Summarize this contract clause in one sentence: ..."}],
)
print(completion.choices[0].message.content)

Four things from Moonshot's quickstart that trip up code written for other models:

1. Send the whole assistant message back. In multi-turn chats and tool loops, append the complete assistant message the API returned, not just content. Reasoning is part of the state.
2. Streaming has two channels. Deltas arrive as reasoning_content and content. Parse structured output from content only.
3. tool_choice="required" works. Unlike Claude Sonnet 5.5, which rejects forced tool use, K3.1 Pro accepts required to force at least one tool call.
4. Video goes through the Files API. Upload with purpose="video" and reference it as ms://<file id> in a video_url content part.

If you're still on K2-series models: Moonshot stops offering them to newly registered users on October 31, 2026 and plans to retire all K2 models by the end of December 2026. Setting up Kimi in an existing coding tool is covered in how to use Kimi K3.

Should you use Kimi K3.1 Pro, Sonnet 5.5 or Opus 5.5?

Our take, for teams building agents and automations:

• Pick Kimi K3.1 Pro for long agent runs where output-token spend dominates the bill and latency isn't critical: research agents, document pipelines, overnight coding tasks. Its token frugality is the real saving. It's also the model to prototype on now if you plan to self-host or need a non-US-hosted option after November 4.
• Pick Claude Sonnet 5.5 when speed matters (it streams nearly twice as fast on Artificial Analysis's numbers), when your workload is cache-heavy (cache reads at $0.10), or when you're already on Anthropic's tool ecosystem.
• Pick Claude Opus 5.5 for the hardest tasks, where three index points and a stronger track record in production agents justify twice the token price.

The gap at the top is now small enough that the right answer depends on your tasks, not the leaderboard. We set up exactly this kind of side-by-side, on your own data, before recommending a model in our AI agent development work. You can also compare prices and context windows across providers in our model directory.

FAQ

What is Kimi K3.1 Pro?

Kimi K3.1 Pro is Moonshot AI's flagship large language model, released on October 23, 2026. It is a post-trained version of Kimi K3 (2.8 trillion total parameters, 104 billion active), with a 1-million-token context window, always-on reasoning, and text, image and video input. Moonshot says its weights will be published by November 4, 2026.

How much does Kimi K3.1 Pro cost?

On the Kimi API it costs $2 per million input tokens, $10 per million output tokens and $0.20 per million cached input tokens. The Highspeed variant costs $4 / $20 with the same $0.20 cached-input price. Access requires a top-up of at least $1.

When will Kimi K3.1 Pro be open weights?

Moonshot says "by Nov 4, 2026." As of October 25, 2026 the weights were not yet published and no license had been announced. The open release has cyber-related training data removed.

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

Not on independent testing. Artificial Analysis scores Opus 5.5 at 58 and K3.1 Pro at 55. Moonshot's own Terminal-Bench 4.0 result puts them within a point (65.7% vs 66.4%). K3.1 Pro costs half as much per token.

What is the difference between Kimi K3.1 Pro and Highspeed?

Moonshot says Kimi K3.1 Pro Highspeed delivers 2-3x the speed of K3.1 Pro. It costs twice as much for input and output ($4 / $20 per million tokens) but keeps the same $0.20 cached-input price. The model ID is kimi-k3.1-pro-highspeed.

Can you turn off reasoning in Kimi K3.1 Pro?

No. Thinking mode is always on. You can set reasoning_effort to low, high or max; the default is max.

What is the context window of Kimi K3.1 Pro?

1,048,576 tokens, the same as every K3-series model on the Kimi API.