Kimi K3.1 Pro: what it is, what it costs, and when it becomes free

Not yet. This is what we checked on 25 October 2026:

• Moonshot's API works on pay-as-you-go. K3.1 Pro is "unlocked after a successful top-up (minimum $1)", according to the official guide. Moonshot's rate-limit page adds that when your total top-ups reach $5 you receive a $5 voucher, and that your total top-up decides your account tier and rate limits.
• OpenRouter lists it as moonshotai/kimi-k3.1-pro at $2 per million input tokens and $10 per million output tokens. There is no free version.
• Hugging Face has no K3.1 Pro weights yet. The newest Moonshot model there is still Kimi K3, published in June 2026.

If you want a free Moonshot model now, use its predecessor. Kimi K3 has open weights and its own free chat with daily limits at kimi.com. In the Arena coding ranking it scores 1655, 37 points below K3.1 Pro.

Moonshot promises the weights "by Nov 4". When they appear, three kinds of free access usually follow within days or weeks: download for anyone with the hardware, free tiers from inference companies that host popular open models, and free daily quotas on aggregators such as OpenRouter. None of these exists yet, and none is guaranteed. We will update this page when they do.

Moonshot has not published the K3.1 Pro license yet. Its predecessor uses the Kimi K3 License, and since K3.1 Pro is "a post-trained version of Kimi K3", that is the most likely starting point. We read the K3 license in full. In plain words:

• You can use, copy, modify, fine-tune, sell and redistribute the model, free of charge, as long as you keep the copyright notice.
• If your company sells access to language models as a service ("Model as a Service") and has more than $20 million in revenue over 12 months, you need a separate agreement with Moonshot before commercial use.
• If a commercial product built on it has more than 100 million monthly users, or more than $20 million in monthly revenue, it must show "Kimi K3" prominently in its interface.
• These two conditions do not apply to internal use inside a company, or to use through Moonshot's own products and certified partners.

For a person, a freelancer or a small company, that is close to a fully open license. Check the K3.1 license on the day it is published: Moonshot can change the terms.

Prices from Moonshot's launch email, per million tokens. Both versions read up to 1,048,576 tokens, and the FAQ confirms there is no higher price for long prompts.

Item	K3.1 Pro	K3.1 Pro Highspeed
Input	$2.00	$4.00
Output	$10.00	$20.00
Cached input	$0.20	$0.20
Cache write, kept 5 minutes	$2.00	$4.00
Cache write, kept 1 hour	$4.00	$8.00

The Decoder reports that these prices are 33 percent lower than K3's. Simon Willison notes that $2 / $10 is the same as GPT-6.1 Sol and half the price of Claude Opus 5.5.

What a request really costs. Some examples with the standard model, rounded:

• A question of 2,000 tokens with an answer of 1,000 tokens: about 1.4 cents.
• A whole book or code base of 500,000 tokens, with a 3,000-token answer: about $1.03 the first time.
• The same 500,000-token text asked again within the cache window: the repeated part costs a tenth, about 10 cents, plus the answer.

Two things raise the bill. K3.1 Pro always thinks before answering, and the thinking is billed as output; with the default effort (max) the hidden reasoning can be longer than the answer. And writing a long prompt into the cache has its own one-time price, shown in the last two rows.

Three independent signals first, then Moonshot's own numbers.

Arena, blind human votes. In the Arena coding ranking (WebDev) of 23 October 2026, where people compare two anonymous answers and vote for the better one, K3.1 Pro is 8th out of 150 models:

#	Model	Arena score	Weights
1	Claude Opus 5.5 (Max)	1807	closed
2	GPT-6 Astra (Max)	1774	closed
3	Claude Sonnet 5.5 (xHigh)	1769	closed
4	GPT-6.1 Sol (Max)	1749	closed
5	Claude Fable 5.1 (Max)	1742	closed
6	Muse Wonder Full (Max)	1726	closed
7	Claude Sonnet 5.5 (High)	1709	closed
8	Kimi K3.1 Pro (Max)	1692	promised for 4 Nov
9	Claude Opus 5 (Max)	1691	closed
10	GPT-6 Sol (Max)	1681	closed
14	DeepSeek-V4.1-Flash (Max), best open model today	1665	open
17	Kimi K3 (Max)	1655	open

If the weights arrive as promised, K3.1 Pro will be the open model with the highest coding score on Arena, 27 points above DeepSeek-V4.1-Flash. See the full coding ranking.

Artificial Analysis. Its Intelligence Index gives K3.1 Pro 55 points, 11 more than K3, as reported by Simon Willison and The Decoder. Only Claude Opus 5.5 (58), Claude Sonnet 5.5 (56) and Muse Wonder Full (56) score higher. The Decoder notes that the open-weights release would shrink the gap between the best open and the best closed model on that index from 12 points to 3. On GDPval-AA, a test of real office work, Artificial Analysis gives it 1,846 Elo, 312 points above K3 and just behind Claude Opus 5.5 (1,866).

Moonshot's own numbers, to read as claims from the maker:

• Terminal-Bench 4.0, real tasks in a computer terminal: 65.7%, against 66.4% for Claude Opus 5.5, 58.2% for GPT-6 Astra and 13% for K3 (reported by The Decoder; on Moonshot's guide this figure is still a blank placeholder).
• BenchCAD, building 3D CAD models: 92.7% with tools and 72.8% without (The Decoder).
• Finance: on Moonshot's internal tests, +7.2 points over K3 in financial modeling, +11.6 points in financial research, and about 4.2 times K3's result in a simulated business that has to be run over a long period.
• Fewer invented facts: on a stress test built from real conversations where users had flagged errors, the share of false claims fell from 23.35% with K3 to 6.85%, and responses with at least one error fell to 46.4% of K3's rate. Moonshot stresses that this is a deliberately hard test, not the error rate of normal use. On launch day The Decoder reported 8.3% for the same test; the guide now says 6.85%.

A note on honesty: at the time of writing, Moonshot's own guide still shows "[xx.x%]" in the place of three benchmark scores. We report only numbers that are written out somewhere we can link to.

Moonshot describes K3.1 Pro as K3 with much more post-training, mostly reinforcement learning (the model tries tasks, gets scored, and learns from the result) across hundreds of domains at once. To make that stable, the company says it rebuilt every layer: the training framework, long-context inference, the sandboxes where the model runs code, and quality checks for the training environments.

Coding. The guide covers quick fixes in real code bases and complete projects, including games built with Unity, Unreal Engine and Godot. The published demos were made "end to end" by the model:

• a mythical palace in Unreal Engine 5.8 with 36 million triangles and a dragon modeled in Blender;
• a 3D action-adventure game in Godot 4.7.2 with five regions, combat, a two-phase boss fight and a day-night cycle;
• a 3D robotic arm in the browser with 2,317 parts and 268 wire routes, checked by 55 automated browser tests.

The most interesting example is about patience, not graphics. In an internal test, K3.1 Pro built a simulator of human walking on top of an open anatomical model with 296 bones and 402 muscles. Over 167 missions and 420 reviews, it took the simulated person from falling within seconds to walking with alternating steps for 11.2 seconds. When the arms still moved badly, it did not just keep tuning the reward: it tested the simulator itself, found 36 errors in the shoulder-muscle calculations, fixed them and trained again.

3D and CAD. It writes and runs scripts for Blender, FreeCAD and similar tools directly, without special connectors, to build parametric objects such as mechanical parts, furniture and buildings. Demos include a turbofan engine in FreeCAD with 2,100 solid parts and moving spools, a five-part 3D-printable lamp, and a cathedral interior with about 3.3 million triangles that runs in the browser.

Finance. Beyond the internal scores above, Moonshot says K3.1 Pro reaches the best results on OfficeQA Pro v2 (questions over real financial documents) and SpreadsheetBench v2 (complete spreadsheet work, including formulas across sheets and finding errors).

Security limits. This is the most unusual section of the guide. Moonshot found that hacking skill grew together with general coding skill, because security material is mixed into normal training data. For the open release, it removed that material from two training stages. On its internal test of real, recently disclosed vulnerabilities (in the Linux kernel, Node.js and others), the version without this cleaning solved 90.5% of tasks; the released model solves 81.0%. Moonshot says the released model stays behind the best paid models in this area on purpose, and argues that open weights help defenders audit code they maintain.

Kimi K3.1 Pro shares its architecture with Kimi K3. From the Kimi K3 model card:

Item	Kimi K3
Total parameters	2.8 trillion
Active for each token	104 billion
Experts	896, of which 16 used for each token, plus 2 shared
Layers	93
Context	1,048,576 tokens
Input	text and images (vision encoder MoonViT-V2, 401M parameters)
Native precision	MXFP4 weights with MXFP8 activations

It is a "mixture of experts": the model is huge, but each word only goes through a small part of it, which keeps it fast for its size. It was also trained directly at 4-bit precision (MXFP4), so the official weights are already compressed and lose little quality from it.

No, and the weights will not change that.

• Size. At 4-bit, Kimi K3 needs about 1,685 GB just for the weights, by our model page estimate. K3.1 Pro has the same size.
• Official hardware. The vLLM recipe for Kimi K3 lists "at least 8x GB300", NVIDIA's newest data-center GPUs, and several machines for real traffic.

That is a server room, not a gaming PC. What the open weights will change is everything around it: other companies can host it, prices can fall, free tiers can appear, and companies can run it on their own servers for privacy. To see the best model that does fit your own GPU, use Run at home.

Moonshot has released a new open model every few months, each time with a longer context and more agent skills. Publication dates on Hugging Face:

• Kimi K2 Thinking: 4 November 2025
• Kimi K2.5: 1 January 2026
• Kimi K2.6: 14 April 2026
• Kimi K2.7 Code: 11 June 2026
• Kimi K3: 13 June 2026
• Kimi K3.1 Pro: API on 23 October 2026, weights promised by 4 November 2026

The K2 family is now being retired: from 31 October 2026, new accounts can no longer use K2 models on the API, and Moonshot plans to switch them off by the end of December 2026.

The API is compatible with the OpenAI SDK. Get a key in the Kimi console after the $1 top-up. The model names are kimi-k3.1-pro and kimi-k3.1-pro-highspeed.

from openai import OpenAI

client = OpenAI(api_key="YOUR_KEY", base_url="https://api.moonshot.ai/v1")
reply = client.chat.completions.create(
    model="kimi-k3.1-pro",
    reasoning_effort="low",  # low, high or max (default)
    messages=[{"role": "user", "content": "Explain Mixture of Experts in two sentences."}],
)
print(reply.choices[0].message.content)

What the official guide documents, and the rules that are easy to miss:

• Thinking is always on. You choose only how much: reasoning_effort accepts low, high and max, and max is the default. Use low for simple tasks to save time and money.
• Answer length. Up to 131,072 tokens by default, up to 1,048,576 with max_completion_tokens.
• Fixed settings. Temperature 1.0, top_p 0.95 and the penalties are fixed: leave them out of the request.
• Streaming sends the reasoning (reasoning_content) and the answer (content) as separate pieces, so you can show or hide the thinking.
• Images and video. Send images as base64, or upload a file and pass its ms:// id; public image links do not work. Video files are uploaded the same way and can be summarized.
• Structured output. response_format with a JSON schema and strict: true forces the answer into your format. Parse content, never the reasoning.
• Partial mode lets you start the answer yourself (for example "Conclusion: ") and have the model continue.
• Tools. Standard function calling, with tool_choice="required" to force a call. In multi-turn conversations, send back the whole assistant message, not just its text, or the model loses its reasoning.
• Dynamic tools. You can add a tool definition halfway through a conversation, inside a system message, and it becomes available from that point.
• Caching is automatic. The cache lasts 5 minutes by default; for workloads with longer pauses, ask for the 1-hour cache with prompt_cache_options.
• Moonshot's own tools (through its "Formula" system) can be plugged in, but its web search "is being updated and is not recommended" for now.