Kimi K3.1 Pro Pricing: $2/$10, a 943K Output Cap, and Weights That Do Not Exist Yet

There is one number on Kimi K3.1 Pro's price card that makes the rest of it strange: the output ceiling. The model is sold at $2.00 per million input tokens and $10.00 per million output, across a 1M-token context window, with a maximum output of 943,000 tokens. That is not a typo and it is not a rounding of the context figure — it is what the serving stack will actually let a single response run to, and it changes what "expensive" means here. A model priced like a mid-tier workhorse is being sold with a long-form capability that most models at this price do not have.

The second strange thing is that Moonshot is selling it as open-weight without having shipped weights. Its own documentation says the full model weights "will be released by Nov 4," and Artificial Analysis — which evaluates what it can call — still labels the model proprietary. Until that date passes, every price quoted for Kimi K3.1 Pro is a hosted price, and there is no self-hosted alternative to compare it against. Both facts belong in a pricing article because both determine what you are actually buying.

The card, line by line

Read off the model's live page on OrcaRouter, 29 October 2026:

• Input — $2.00 per million tokens.

• Output — $10.00 per million tokens.

• Context — 1M tokens, with a 943K maximum output.

• Modalities — input text, image and video; output text.

• Endpoints — /v1/chat/completions, /v1/responses, and /v1/messages.

• Measured latency — p50 time to first token 7.26 seconds; p95 10.00 seconds, over seven days.

• Traffic — 2,779.4 million tokens over the same seven days.

That last pair is the part most price pages omit, and it is the most useful thing on the page: 2.78 billion tokens in a week is evidence that the endpoint is being used at scale, not that it is fast. The latency figure says the opposite — 7.26 seconds to first token is slow, and it is slow in a way that matters differently depending on what you are building.

What the $10 output actually buys

Per-token pricing is a poor unit for planning, so here is the same card expressed as tasks. A single request that reads 200,000 tokens of context and writes 5,000 tokens costs $0.40 in and $0.05 out — about 45 cents. That same request against GPT-6.1 Sol, which carries the same $2.00 / $10.00 card, costs the same 45 cents and scores three Index points lower on the independent board. That is the case for Kimi K3.1 Pro in one sentence, and it is also the case against relying on the card alone: identical prices mean the decision is entirely about which model does the work.

Now the other end of the range, which is where the 943K output cap stops being a curiosity. An agent that generates a long document or a large code file in one call can plausibly write 100,000 output tokens. At $10 per million that is $1.00 for the response — five times the cost of the input side if the prompt was small. Output tokens are where this card gets expensive, and long-form generation is where a 943K ceiling will actually be exercised.

Two honest gaps in what the page tells you: it lists input and output rates and no cache-read or batch rate, so there is no published discount to plan around; and it gives no per-account rate limit, which means the ceiling you hit under load is an operational discovery rather than a budget line. Both are worth asking about before you move a high-volume job onto it.

Latency is the second bill

7.26 seconds to first token does not appear on the rate card, and for a chat interface it is almost irrelevant — the reader is waiting anyway. For an agent loop it is a different number entirely. A loop that takes twenty turns pays roughly two and a half minutes of pure waiting per task at p50, and closer to three and a half at p95, before any of the actual work is counted. Multiply by parallel agents and the wall clock, not the invoice, becomes the constraint.

That is the specific trade this model asks you to make: priced identically to a faster mid-tier model, scoring higher on a composite index, and slower at the front of every request. The right way to decide is to measure your own p95 on your own prompts rather than trusting either vendor's figure.

The endpoint nobody mentions, and the one that decides adoption

The model page lists three endpoints, and the third is the interesting one. Beyond the OpenAI-shaped chat completions and responses routes, there is a /v1/messages endpoint — the Anthropic-shaped surface. For anyone running a coding harness that speaks Anthropic's API, that is the difference between a model you can point at your existing tooling and a model that needs an adapter.

It also reframes the price. A harness like Claude Code is expensive to run not because of any single call but because of how many calls a session makes. At $2.00 / $10.00 with a 1M window, the arithmetic that usually kills these setups — a long context being re-sent on every turn — is the one this card is best shaped for, provided the 7-second first token does not make the session feel broken.

4 November, and what it does to the price

The open-weights date is a pricing event, not just a licensing one, and there is a precedent for how it plays out. Kimi K3 was announced on 16 July 2026 and its weights appeared on Hugging Face on 27 July — an eleven-day gap. If K3.1 Pro follows the same shape, the hosted price stops being the only price in the market: self-hosting becomes a real option for anyone with the GPUs, and the hosted price has to compete with a marginal cost that is somebody else's electricity bill.

What that does not change is this: self-hosting a model this size is a capital decision, not a discount. The hosted card is the cheap way to use it, and the weights are the cheap way to own it — different buyers, different answers. Anyone planning a large commit should treat 4 November as the date the comparison set changes, not as a date the price drops.

How to budget it without guessing

Three practical rules, in order of importance.

• Budget output first. At a 10:1 output-to-input price ratio, the output side dominates every workload that generates rather than reads. If you do not know your output-token share, measure it before you model the cost.

• Treat 1M context as a billing hazard, not a feature. A large window invites you to send everything; sending everything on every turn is how a cheap model becomes an expensive one. Cache behaviour is undocumented on this page, so assume every token is charged every time until you can measure otherwise.

• Price the alternative on the same day. Rates move. On OrcaRouter the provider's published price is what you are billed, with no markup, so a Moonshot price change is live here the same day it is announced — which also means a comparison you wrote down last month may be describing a card that no longer exists.

What to watch: whether the weights land on 4 November and in what licence, whether the model page gains a cache-read rate once open-weight serving makes batching normal, and whether the hosted card moves at all in response. The price you pay for Kimi K3.1 Pro this week is a hosted price with an expiry-shaped question attached to it.