Kimi K3.1 Pro Providers: 17 Endpoints Priced and Timed

Moonshot AI opened the weights of Kimi K3.1 Pro on 4 November 2026. A week earlier, when we first tested Kimi K3.1 Pro pricing, Moonshot was the only host on OpenRouter. Today OpenRouter lists 20 endpoints from 15 companies, all serving the same model, with input rates from $0.39 to $4.00 per million tokens. So the useful question has changed from what K3.1 Pro costs to which provider you should buy it from.

We sent the same support-bot request from our WordPress server to 17 of those endpoints, pinned one at a time, and priced and timed every answer. We also ran OpenRouter’s default routing and its cheapest and fastest shortcuts, plus Claude Sonnet 5.5 and Claude Haiku 5.5 for scale. That came to 184 requests.

The short answer: at identical token counts, 1,000 answers cost $6.22 on Relace and $28.28 on Moonshot’s Highspeed endpoint, a gap of 4.5 times, and 12 of the 17 endpoints sit between $12.72 and $14.91. Speed barely follows price: PrimaLabs took a median 3.54 seconds at $1.80 / $9.00, while Moonshot’s own endpoint took 21.24 seconds at $2.00 / $10.00. Every answer that came back at default settings carried the core facts on every host. If cost matters and you don’t need K3.1 Pro specifically, Claude Haiku 5.5 did the same job for $0.85 per 1,000 answers.

Kimi K3.1 Pro providers and rates

These are the rates on OpenRouter’s endpoints API on 5 November 2026, per million tokens. We left out three duplicates: US-only routes from Sail Research and InferenceNet, which list the same rates as the routes we tested.

Relace is the outlier on input, not output. Its input rate is about 80% below Moonshot’s, while its output rate is only about 21% lower. That matters because a support bot sends a lot of input: the system prompt and knowledge base passages arrive with every question. Sail Research has the opposite shape, with cheap input and output at three times the usual rate, and on a reasoning model that is a bad trade. The rates are as listed, and prices on a newly hosted model can move quickly, so check them again before you commit.

How we tested it

Each request was the one a MxChat support bot sends: our site’s real system prompt plus four passages from our own documentation, about 4,174 tokens on Kimi’s count, followed by one visitor question. We asked four questions, each twice:

1. Does MxChat work with WooCommerce, and can the bot look up an order status for a logged-in customer?
2. Do I need Pinecone to use the knowledge base, or can I store embeddings in WordPress?
3. Can a visitor be handed over to a human on Slack or Telegram when the bot cannot answer?
4. Can I use my own OpenRouter or OpenAI API key, and which models can I choose from?

Each endpoint was pinned with no fallbacks and reasoning left at its default. Then we added Relace with reasoning off, Sonnet 5.5, Haiku 5.5, and three unpinned runs: the plain model ID, :floor and :nitro.

One thing makes this harder than reading a rate card. K3.1 Pro thinks before it answers, and the length of that thinking varies a lot between identical requests. Across our 135 pinned K3.1 Pro answers, output ranged from 98 to 1,508 tokens, about 89% of it reasoning. With only eight answers per host, a host can look cheap simply because the model happened to think less there. So the first cost column below prices every host at the same tokens: 4,174 input and 579 output, the mean across those answers. That is the column to compare hosts on. The measured column is what each host’s eight answers actually came to.

What 1,000 answers cost on each host

For a site handling 10,000 questions a month with no cache, that is about $62 on Relace, about $141 at the common $2.00 / $10.00 rate, and about $283 on Moonshot Highspeed.

Most of the market has settled on one price. Seven endpoints list exactly Moonshot’s $2.00 / $10.00, and the 10% cheaper group at $1.80 / $9.00 saves about $1.42 per 1,000 answers. Picking among those on cost alone would be a waste of effort. Picking on speed isn’t, as the next section shows.

The cache helps, but unevenly. On the second pass, Modal and PrimaLabs read about 4,100 of 4,174 input tokens from cache and Moonshot about 4,030, while Wafer read about 1,630 and Phala about 1,400. That is why Moonshot’s warm figure of $6.63 comes close to Relace’s $4.93 despite the very different input rates. Relace’s cached rate of $0.20 is only about half its input rate, so its warm saving is small. The warm column also carries the same reasoning noise as the measured column, and eight answers can’t tell you which host caches best over a whole day. Treat it as a hint.

Speed: price doesn’t buy it

The fastest endpoint was InferenceNet Fast at a median 2.77 seconds, for 1.67 times the price of InferenceNet’s standard route, which took 4.23 seconds. PrimaLabs (3.54 s) and Relace (3.60 s) were nearly as quick at far lower prices.

The surprise is Moonshot itself. Its standard endpoint took 21.24 seconds, slower than the 16.18 seconds we measured on 29 October, and only Nebius (32.09 s) was slower. Moonshot’s Highspeed route charges double and took 13.13 seconds. Thirteen of the other endpoints beat Highspeed, most at about half its price. We can’t see each host’s load from here, and the day after an open-weights release is an unusual day, so this is a snapshot. But it is a large difference on identical requests.

The slowest answer matters more than the median on a live widget. DeepInfra’s median was 12.15 seconds, but two of its eight answers took over a minute, the worst at 102.21 seconds. Nebius had one at 64.54 seconds, and Parasail, with a 6.91-second median, had one at 53.86 seconds. On the other 14 endpoints, no answer took longer than 31 seconds.

Reliability was otherwise good. Every request to 16 of the 17 endpoints succeeded on the first attempt. Sail Research was the exception: three requests needed retries, and a fourth failed with HTTP 429 after four attempts, which is why it has seven answers.

Does the host change the answers?

We didn’t see it. Listed formats differ (fp4, mxfp4, fp8, bf16), and Moonshot’s model card describes quantization-aware training for MXFP4, so differences are plausible. In practice, all 135 pinned answers at default settings passed our check for the core facts: order lookups need Pro and a logged-in customer, Pinecone is optional, Slack and Telegram handoff exist, and you can bring your own key. Answers averaged 36 to 42 words on every host. Our check looks for facts in short answers, though. It wouldn’t detect a host doing worse on long agent work or hard reasoning, so don’t read this as proof the formats are equivalent.

Relace with reasoning off

The cheapest host gets cheaper. Sending "reasoning": {"enabled": false} on Relace cut the measured cost from $5.86 to $2.29 per 1,000 answers and the median from 3.60 to 1.23 seconds, and warm billing was $1.90. Seven of eight answers were right. On the eighth, the Slack and Telegram question, it declined and pointed the visitor to the documentation, although the answer was in the passages it had been given. Eight answers can’t show whether that is a pattern, but it was the only miss among all the K3.1 Pro answers we collected. Check your own questions if you go this way.

Is K3.1 Pro the right model for a support bot at all?

K3.1 Pro on Relace costs about 64% less than Claude Sonnet 5.5 at $16.45, and it’s only slightly slower (3.60 s against 2.79 s). Part of that comes from the tokenizers: Claude counts the same request as 6,838 input tokens where Kimi counts 4,174. If you were paying for Sonnet-class answers, that is a real saving.

But a four-passage support question isn’t what a 2.8-trillion-parameter model is for. Claude Haiku 5.5 answered all eight correctly at $0.85 per 1,000 answers with a 1.88-second median. That is about 7 times cheaper than K3.1 Pro on Relace with default reasoning, and 2.7 times cheaper than with reasoning off. K3.1 Pro may earn its cost with a very large context, up to 1,048,576 tokens, or on long multi-step work. We didn’t test either here.

What happens when you don’t pick a host

Most WordPress site owners won’t pin anything. MxChat, for example, sends OpenRouter a model ID and no provider preference. We ran that case too, along with the two shortcuts OpenRouter documents, eight answers each:

Default routing doesn’t find you the cheapest host. It spread eight requests across five companies and cost about the same as the common $2.00 / $10.00 rate. Adding :floor cut the bill by more than half, with no other change. These runs came after the main test, so caches were partly warm and the amounts actually billed were lower than the table shows. The listed-rate column is the fair comparison. And :nitro sorts by throughput, not by time to a finished answer, so it isn’t guaranteed to pick the fastest host from our table.

Choosing a Kimi K3.1 Pro provider in MxChat

MxChat’s OpenRouter picker loads models from OpenRouter’s model list, which doesn’t include :floor or :nitro variants, and it has no host setting. There is a way around that on OpenRouter’s side:

1. Create an API key on OpenRouter and add credit. Our documentation shows where the key goes in MxChat.
2. Select moonshotai/kimi-k3.1-pro in MxChat’s OpenRouter model picker, or Haiku 5.5 if the section above changed your mind.
3. In your OpenRouter account preferences, set the default provider sort to price. OpenRouter’s documentation says this replaces its default routing. It’s account-wide, so it also affects other apps using that account; a separate account avoids that. We tested :floor, not this setting, so check which host your first few requests go to in your OpenRouter activity.
4. If you see a host stalling, OpenRouter also has an account-wide list of ignored providers.
5. Send your bot a few real visitor questions and check the answers before you switch it on for everyone.

One limit to know about: MxChat stops an OpenRouter request after 60 seconds and doesn’t send a reasoning setting. Three of our 135 pinned answers ran longer than that (two on DeepInfra, one on Nebius), without streaming. Under MxChat those requests would hit the limit, so sorting away from those two hosts is worthwhile. K3.1 Pro is nothing like Tencent Hy4, whose default reasoning ran past 60 seconds on nearly every answer. Order lookups for logged-in WooCommerce customers need MxChat Pro.

What the open weights change, and the licence

For almost every WordPress site, the weights matter because of the competition they create, not because you would run them. The release is 96 weight files totalling about 1.56 TB, for a model with 2.8 trillion parameters, of which 104 billion are active. Moonshot recommends vLLM, SGLang or TokenSpeed to serve it. That is a job for a GPU cluster rather than a web host. Once several companies can serve a model, prices spread, as we saw on Kimi K3, GLM-5.3 and Nemotron 3.5 Super.

The licence isn’t a standard open-source one, though it’s permissive. We read it in the repository, and in summary:

• Use, modification, redistribution and commercial use are allowed, with the copyright notice kept.
• A company running a business selling API access to models, with group revenue above $20 million over 12 months, needs a separate agreement with Moonshot.
• A commercial product over 100 million monthly users or $20 million in monthly revenue has to display the model’s name prominently.
• Those two conditions don’t apply to internal use, or to use through Moonshot’s own products or certified inference partners.

A site calling K3.1 Pro through a host’s API for its own chatbot is very unlikely to reach any of those conditions. They mainly concern the hosts. This is our reading rather than legal advice, so read the licence itself if you are near those numbers or plan to resell access.

If you’re weighing OpenRouter against direct billing, our OpenRouter vs direct API guide covers what the router adds.

Frequently asked questions

Which Kimi K3.1 Pro provider is cheapest?

Relace, at $0.39 input and $7.94 output per million tokens on OpenRouter on 5 November 2026. At the same token counts, that is $6.22 per 1,000 chatbot answers, against $14.14 at the common $2.00 / $10.00 rate.

Which Kimi K3.1 Pro provider is fastest?

In our test, InferenceNet Fast at a median 2.77 seconds, followed by PrimaLabs at 3.54 and Relace at 3.60. That was eight answers each on one day, so check it again before you rely on it.

Is Moonshot’s own API the best place to run it?

Not on price or speed in this test. Its standard endpoint took 21.24 seconds at $2.00 / $10.00, and Highspeed took 13.13 seconds at $4.00 / $20.00. It did cache well, with a warm cost of $6.63 per 1,000 answers.

Are Kimi K3.1 Pro’s weights really open?

They are public and ungated on Hugging Face under a custom licence, not a standard open-source one. Commercial use is allowed, with conditions for very large API providers and very large products.

How do I get the cheapest host on OpenRouter?

In your own code, add :floor to the model ID, or send a provider sort by price. Without code, set the default provider sort to price in your OpenRouter account preferences, which covers every request on that account.

Does the host change answer quality?

Not on our four short support questions: all 135 default answers carried the core facts. Our check doesn’t cover harder work such as long agent tasks.

Can I use Kimi K3.1 Pro in MxChat?

Yes. It’s in OpenRouter’s model list, and nearly all our answers finished well inside MxChat’s 60-second limit. MxChat can’t choose a host or turn reasoning off, so use the OpenRouter account settings above.

Method note: tests run 5 November 2026 from mxchat.ai’s server in the US through OpenRouter, 4 questions × 2 passes per setup, temperature 1, without streaming. Rates were read from OpenRouter the same day. “Same tokens” uses each endpoint’s listed rates at 4,174 input and 579 output tokens. “Measured” uses each setup’s actual tokens at listed rates. “Warm” is the mean OpenRouter billed on the second pass, and setups for the same model on one host may share a cache. On Relace, reported reasoning tokens sometimes exceeded total output tokens, so its token accounting differs slightly from other hosts; billed amounts stayed close to our calculation. Your costs depend on your prompt size, answer length and traffic pattern.