Kimi K3.1 Pro vs Claude Opus 5.5: Pricing, Benchmarks, and What a Coding Task Really Costs

Moonshot AI released Kimi K3.1 Pro on October 23, 2026, calling it its “most capable 3T-class open-weight model.” It is the first model of the K3.1 series: 2.8 trillion parameters with 104 billion active, a 1 million token context, and image and video input. It is live on the Kimi API and OpenRouter today, and Moonshot says the weights will follow by November 4.

On the API it costs $2 per million input tokens and $10 per million output, a third less than Kimi K3 ($3 and $15) and half the list price of Claude Opus 5.5 ($4 and $20).

Moonshot’s own headline comparison is with Opus 5.5: on Terminal-Bench 4.0, K3.1 Pro scores 65.7%, against 66.4% for Opus 5.5 and 58.2% for GPT-6 Astra. So we ran both on the same coding task in the same agent, three runs each, and K3.1 Pro at two reasoning effort levels.

Model IDs. Kimi API: kimi-k3.1-pro and kimi-k3.1-pro-highspeed, base URL https://api.moonshot.ai/v1. OpenRouter: moonshotai/kimi-k3.1-pro. Weights: not on Hugging Face yet. In Atomic Agent, use the Moonshot AI (Kimi) preset and set temperature to 1 (details below).

What is Kimi K3.1 Pro?

Kimi K3.1 Pro is Moonshot AI’s October 2026 upgrade of Kimi K3, its open-weight flagship from July. It keeps K3’s scale (2.8T total, 104B active parameters, mixture of experts) and 1M token context, and adds what Moonshot describes as stronger long-horizon coding, 3D modeling in Blender and CAD, and finance work, with fewer hallucinations.

From Moonshot’s launch post, tech blog, K3.1 Pro quickstart, and pricing page, October 23, 2026.

There is also K3.1 Pro Highspeed, which Moonshot says runs 2 to 3 times faster and costs twice as much ($4 input, $20 output). We did not test it.

How much does Kimi K3.1 Pro cost, and is it worth it?

Kimi K3.1 Pro costs $2 per million input tokens and $10 per million output on the Kimi API, half of Opus 5.5’s list price. In our coding agent test, that became about $0.16 to $0.21 per task at low effort and $0.35 to $0.46 at the default max effort, against $0.37 for Opus 5.5. Low effort is where the savings show up.

List prices from Moonshot and OpenRouter on October 23, 2026. Moonshot also charges for 5 minute cache writes at the input price ($2 per million for K3.1 Pro).

The API is pay as you go. K3.1 Pro unlocks after a top-up of at least $1, and the total you have topped up sets your rate limits.

How we tested

We used the same task as in our Claude Sonnet 5.5 test and our DeepSeek V4.1 Flash 1017 test: a small Python package with five planted bugs and two missing features (a settlement function and a command-line tool), plus 32 hidden tests the agent never sees. The prompt is one line: read the task file, complete it, run the tests until they pass. Each setup ran three times in Atomic Agent v0.6.9 with approvals off.

Opus 5.5 ran through OpenRouter, pinned to Anthropic. K3.1 Pro was not on OpenRouter yet when we started, so it ran directly on the Kimi API with temperature set to 1, the only value the model accepts. We ran it twice: once with Moonshot’s default reasoning effort (max) and once with reasoning_effort: "low". All runs were on October 23, 2026, within about an hour of launch.

Results

Three runs per setup, averages shown. Opus, Sonnet, DeepSeek, and GLM costs are what OpenRouter billed; the Sonnet, DeepSeek, and GLM rows are from our October 19 runs on the same task and agent. Kimi costs are a range: the low end is the token counts from the agent’s logs at Moonshot’s list prices, and the high end adds a third, because our Kimi balance dropped $2.02 across all six Kimi runs while the token math came to $1.52. We could not split that difference between the two effort levels.

Every setup solved the task, so the table is not a ranking of intelligence. It shows how much work each model does to get there, and that work is what you pay for.

K3.1 Pro works in many small steps. At max effort it averaged 28 model calls per run against 7 for Opus, and wrote 12,000 tokens of reasoning per run against Opus’s 700. Each call is cheap, and about 96% of its input was served from cache at $0.20 per million, but the extra calls add up. Turning reasoning down to low cut the reasoning tokens by almost 90% and the steps by a third, and that is what brought the cost per task under Opus.

Speed is the trade-off. Even at low effort, K3.1 Pro took about twice as long as Opus on this task, mostly because it makes more calls. The Highspeed variant is Moonshot’s answer to that; we have not measured it yet.

Limits of this test. One small, well-specified Python task, three runs per setup, on launch day, when the API was likely under heavy load. It shows cost and speed on a task every model can solve. It says nothing about the long-horizon, 3D, and finance work Moonshot highlights, or about the Highspeed variant.

Download Atomic Agent to repeat this on your own repo: switch the model per run and keep the same tools and prompt.

How does Kimi K3.1 Pro score on benchmarks?

Moonshot’s numbers put K3.1 Pro next to the top closed models rather than ahead of them. On Terminal-Bench 4.0 it scores 65.7%, just behind Claude Opus 5.5 (66.4%) and well ahead of GPT-6 Astra (58.2%). On Moonshot’s own long-horizon coding benchmark the order flips: K3.1 Pro beats Opus 5.5 but trails GPT-6 Astra. It leads on OfficeQA Pro v2 (64.4) and on Moonshot’s hallucination test among the open models it lists.

Vendor-published numbers from Moonshot’s tech blog and launch thread. Kimi Code Bench and Kimi Hallu Bench are Moonshot’s internal evaluations. We did not re-run any of them. Moonshot used max reasoning effort for its evaluations.

Is Kimi K3.1 Pro free?

No, not on the API. Moonshot unlocks K3.1 Pro after a top-up of at least $1 and bills per token from there. The model is also live in Kimi.ai, Kimi Work, and Kimi Code, which have their own subscriptions. OpenRouter lists it only as a paid model served by Moonshot, with no free route.

Kimi K3.1 Pro vs Claude Opus 5.5: which should you use?

On our task, Kimi K3.1 Pro at low effort did the same job as Opus 5.5 for about half the money, in about twice the time. At its default max effort it lost that edge.

• Budget agent work you can leave running: K3.1 Pro with reasoning_effort: "low". Same result on our task at about half the cost per task.
• Interactive work where you wait for each answer: Opus 5.5, or try K3.1 Pro Highspeed. Opus finished in about a minute; K3.1 Pro took two to five.
• Open weights, self-hosting, or avoiding one vendor: K3.1 Pro, once the weights land by November 4. Opus has no open weights.

How do I use Kimi K3.1 Pro via API, Claude Code, or Atomic Agent?

Any agent that speaks the OpenAI or Anthropic API can use it. The Kimi API’s OpenAI-compatible base URL is https://api.moonshot.ai/v1, and its Anthropic-compatible one, for Claude Code and similar tools, is https://api.moonshot.ai/anthropic. Use kimi-k3.1-pro as the model name. Two things to know: the model only accepts temperature: 1, and reasoning effort is a top-level reasoning_effort field with low, high, or max.

Install and start Atomic Agent:

curl -fsSL https://atomicagent.io/install | sh
atomic-agent

On Windows, run irm https://atomicagent.io/install.ps1 | iex in PowerShell, or download the desktop app.

In the provider wizard, pick the Moonshot AI (Kimi) preset and paste your key from platform.kimi.ai. It saves the key as MOONSHOT_API_KEY, and kimi-k3.1-pro shows up in the model list because the list comes from Moonshot’s server.

Then add the Kimi settings to the moonshot entry under llm.providers in ~/.atomic-agent/config.json. The entry should look like this:

{
  "id": "moonshot",
  "kind": "openai-compatible",
  "baseUrl": "https://api.moonshot.ai",
  "apiKeyEnvVar": "MOONSHOT_API_KEY",
  "defaultChatModel": "kimi-k3.1-pro",
  "extraBody": { "temperature": 1, "reasoning_effort": "low" }
}

Without the temperature line, the current Atomic Agent sends its default temperature, Moonshot rejects the request, and the agent falls back to your next provider. We found this while testing and are fixing it so K3.1 Pro works without the extra line. Drop reasoning_effort to use Moonshot’s default max. More on providers in our models guide.

You can also pick it on OpenRouter as moonshotai/kimi-k3.1-pro, with one key for every model in the table above; see our OpenRouter guide. The temperature note applies there too.

Can you run Kimi K3.1 Pro locally?

Not on a normal computer. With 2.8 trillion parameters, even a heavily quantized build needs hundreds of gigabytes of memory, which means a server with several data-center GPUs. The weights are not out yet either; Moonshot promises them by November 4, 2026. Expect vLLM and SGLang support first and community quantizations after that.

If you want an agent that keeps most of the work on your own hardware, there are two workable paths:

• Use K3.1 Pro as the planner only. Atomic Agent’s Fusion mode can use a cloud model as the orchestrator, which plans and reviews, while local workers do the edits. Only the orchestrator calls the API, though it still reads your project files.
• Pick a model that fits your machine. We ran a model that actually runs on your laptop through the same kind of agent work.

Is K3.1 Pro better than Kimi K3?

By Moonshot’s numbers, yes, at a lower price. It reports 61.2% against K3’s 56.0% on its long-horizon coding benchmark, 7.2 points better on financial modeling, 11.6 points better on financial research, and a claim-level hallucination rate on Kimi Hallu Bench Hard that fell from 23.4% for K3 to 8.3%. It also costs a third less per token. We did not run K3 in this test.

Was Kibitz on Arena Kimi K3.1?

Probably. Moonshot has not confirmed the codename, but the model showed up in public well before launch:

• Late September: developers spotted a kimi-k3-1 identifier on Moonshot’s API platform.
• October 21: Kimi K3.1 appeared in the model switcher config of Kimi Code.
• October 22: an anonymous model called Kibitz started showing up on Arena, and testers posted 3D scenes and one-shot websites from it, the same strengths Moonshot leads with. Moonshot tested Kimi K3 the same way under the name Kivine.
• October 23, morning: Moonshot’s Kimi account posted a short video of Kimi Code with “Model: Kimi K3.1” on screen.
• October 23, 20:15 UTC+3: release, as Kimi K3.1 Pro.

Is Kimi K3.1 Pro open source?

Moonshot calls K3.1 Pro open-weight and says the full weights will be released by November 4, 2026; on launch day they were not on Hugging Face. Kimi K3 is published under the Kimi K3 License, an MIT-style license with two extra conditions: a company running a Model as a Service business with more than $20 million in yearly revenue must sign a separate agreement with Moonshot, and products with more than 100 million monthly users or $20 million a month in revenue must show “Kimi K3” in their interface. Moonshot has not said yet whether K3.1 Pro uses the same license. It also says it removed cyber-related data from K3.1 Pro’s training before the open-weight release.

If you want to compare Kimi K3.1 Pro with other models on your own code, download Atomic Agent and run the same task with /model. For more options, see our list of open-source alternatives to Claude Code.