Kimi K3.1 Pro Pricing: Tested on a WordPress Chatbot

Kimi K3.1 Pro pricing is $2.00 per million input tokens and $10.00 per million output tokens, with cached input at $0.20. Moonshot AI released the model through its API on 23 October 2026 and says the open weights will follow on 4 November. Those are exactly the list rates of Claude Sonnet 5.5, and a third lower than Moonshot’s own price for Kimi K3 ($3 / $15). A Highspeed tier costs double: $4 / $20.

Sticker prices only tell you so much. We wanted the number a WordPress site owner pays: what 1,000 chatbot answers cost on K3.1 Pro, and how long a visitor waits for each one. So on 29 October 2026 we sent the same retrieval-augmented support-bot request from our own WordPress server to K3.1 Pro at five settings, to Kimi K3 on three different hosts, and to four rival models. That was 96 calls. All of them returned HTTP 200 and all 96 answers were correct.

The short answer: 1,000 answers cost $12.76 on Kimi K3.1 Pro with no cache, against $17.17 on Claude Sonnet 5.5 at the same per-token rates. Kimi is 26% cheaper per answer because its tokenizer counted our prompt as 4,174 tokens where Claude counted 6,838. The catch is speed. K3.1 Pro took a 16.2-second median to answer, against 3.7 seconds on Sonnet 5.5. With reasoning switched off it answered in 6.8 seconds for $8.90 per 1,000, and every answer was still correct.

Kimi K3.1 Pro pricing at a glance

These are the rates on OpenRouter’s model API on 29 October 2026, which match the figures Moonshot published at launch. On OpenRouter, Moonshot is the only host for K3.1 Pro so far. It offers two endpoints, standard and Highspeed, and lists both as MXFP4-quantized.

The Kimi K3 rows matter because of what happens next. Kimi K3 is an open-weight model, and OpenRouter now lists 31 endpoints for it, from Relace at $0.18 per million input tokens up to $4.50 on Fireworks’ fast tier. When K3.1 Pro’s weights ship on 4 November, the same thing will probably happen to it. Today, though, Moonshot is the only place to buy it. (Our Kimi K3 hosts test covers K3’s host spread in detail.)

How we tested it

Every request was the one a MxChat support bot actually sends. It contained the site’s real system prompt plus four passages pulled from our own documentation, about 4,100 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?

We ran 12 setups:

• Kimi K3.1 Pro on Moonshot’s standard endpoint at its default setting, at reasoning_effort low and high, and with reasoning switched off.
• Kimi K3.1 Pro Highspeed at its default setting.
• Kimi K3 on Moonshot, Relace and Fireworks, each provider pinned with fallbacks disabled.
• Claude Sonnet 5.5 (same list price as K3.1 Pro), Claude Haiku 5.5, GPT-6.1 Terra at low effort, and Mercury 3.

All calls went through OpenRouter from our WordPress server in the US, non-streamed, with usage accounting switched on so OpenRouter reported the exact tokens and billed cost of each call. The first pass of each setup is the “cold” cost; the second, identical pass shows what provider caching does. We graded each answer against the facts in the passages: the WooCommerce answer had to say the feature needs MxChat Pro and a logged-in customer, the Pinecone answer had to say WordPress storage works, and so on.

What 1,000 chatbot answers cost

Four things stand out.

Same rates as Sonnet 5.5, a quarter off the bill. K3.1 Pro and Sonnet 5.5 both charge $2 in and $10 out, and K3.1 Pro actually wrote more output tokens per answer (441 against 349). It still came out $4.41 cheaper per 1,000 answers. The reason is the tokenizer: the identical request was 4,174 input tokens on Kimi and 6,838 on Claude, 64% more. On a retrieval chatbot, where input is most of the bill, that difference beats everything else. We measured the same effect on Claude’s side in our Claude tokenizer test.

K3.1 Pro is 35% cheaper per answer than K3 on Moonshot. $12.76 against $19.54, which tracks the one-third cut in list price almost exactly. Both models used the same 4,174 input tokens and similar output (441 against 468 tokens), so this is a straight price cut, not a change in token usage.

The cheapest Kimi is still the old one on a third-party host. Kimi K3 on Relace cost $4.71 per 1,000 answers, a third of K3.1 Pro, because Relace charges $0.18 per million input tokens against Moonshot’s $2.00. On Relace, 81% of the bill was output. That is the price pressure K3.1 Pro will face once its weights are public.

Kimi is not a budget model. Claude Haiku 5.5 answered the same eight questions correctly for $0.88 per 1,000, and Mercury 3 for $0.48. On our simple support questions, K3.1 Pro bought nothing those models did not. Its pitch is long-horizon coding and agent work, not short answers from a knowledge base.

Speed: the real cost of K3.1 Pro

A chatbot visitor feels time, not tokens. At its default setting, K3.1 Pro took 16.18 seconds at the median to return a 34-word answer, and its slowest call took 28.7 seconds. Sonnet 5.5 took 3.75 seconds. That is more than four times longer for an answer of the same quality.

Two measurements explain it. Moonshot’s standard endpoint delivered about 23 output tokens per second end to end, against 91 on Sonnet 5.5 and 168 on Haiku 5.5. And K3.1 Pro spent most of those tokens thinking: 374 of its 441 output tokens per answer were reasoning the visitor never sees.

The host matters as much as the model. Kimi K3 took 12.98 seconds on Moonshot but 4.47 seconds on Relace and 5.12 on Fireworks, with nearly identical output. If K3.1 Pro follows K3 onto third-party hosts, expect the same: faster answers somewhere other than Moonshot.

Highspeed: twice the price, not twice the speed

Moonshot sells Highspeed as two to three times faster. On raw throughput that held up: 55 output tokens per second against 23, about 2.4 times faster. But Highspeed also reasoned far more on the same questions, 684 reasoning tokens per answer against 374. Faster tokens, nearly twice as many of them. The median answer came back in 13.63 seconds, only 16% faster than standard, and cost $31.85 per 1,000 answers, 2.5 times as much. For a chatbot, Highspeed was the worst-value setup in this test.

Reasoning settings: switch it off for a chatbot

Some launch coverage said K3.1 Pro always reasons and that thinking cannot be turned off. Through OpenRouter, it can. Sending reasoning: {"enabled": false} returned answers with zero reasoning tokens and no reasoning text, on all eight calls. Those answers were all correct, slightly longer (41 words against 34), cost 30% less ($8.90 against $12.76) and arrived in 6.77 seconds instead of 16.18.

One oddity: the default setting reasoned more than explicit “high” effort. With eight calls per setup and wide spread between calls (default ranged from 11.5 to 28.7 seconds), we would not read much into the order of high and default. We would read a lot into “off”, which was faster and cheaper on every single call.

Even with reasoning off, K3.1 Pro was the slowest setup in this test that did no reasoning. Its throughput with reasoning off was about 10 tokens per second, which suggests most of those 6.8 seconds were spent reading the 4,100-token prompt, not writing the answer.

Prompt caching on Kimi K3.1 Pro

Moonshot caches automatically. On our second, identical pass, about 4,030 of 4,174 input tokens came back as cached, billed at $0.20 per million instead of $2.00. That cut the default setup from $12.76 to $4.90 per 1,000 answers, and the reasoning-off setup from $8.90 to $1.96.

That matters for a chatbot, because every conversation repeats the same system prompt and often the same knowledge-base passages. Two caveats:

• Claude cached nothing on these calls. Anthropic’s caching needs explicit cache breakpoints in the request, and our request did not set any, so Sonnet 5.5 cost the same $17.17 on the repeat pass. With breakpoints set, Sonnet’s cached input is $0.10 per million, half of Kimi’s. Our Sonnet 5.5 prompt caching test covers the cache-write cost that comes with it.
• Cheap hosts may not discount the cache at all. Relace cached about 2,940 tokens on the repeat pass of Kimi K3 but charges $0.18 for cached input against $0.1824 for fresh input, so the bill did not move ($4.71 to $4.70).

What it costs per month

Here is what the measured costs come to at three traffic levels, with no cache savings:

With Moonshot’s automatic cache on a repeating prompt, the first two columns fall to roughly $4.90 and $1.96 per 1,000. For a store answering 10,000 questions a month, K3.1 Pro with reasoning off and a warm cache works out to about $20. That is cheap enough. The 6.8-second wait is the harder sell.

Where Kimi K3.1 Pro falls short for a chatbot

• Speed. Even the fastest K3.1 Pro setup (6.77 seconds) was nearly twice as slow as Sonnet 5.5 and almost nine times slower than Mercury 3. Visitors notice a wait of more than a few seconds.
• One host, for now. Until the weights ship, Moonshot is the only provider. If Moonshot has an outage, there is no second route for the same model.
• Variance. Default-setting answers ranged from 11.5 to 28.7 seconds. A chatbot needs consistent waits, not just a good median.
• A small test. Four questions and eight calls per setup show cost, speed and basic accuracy on a retrieval bot. They do not test the long-horizon agent and coding work K3.1 Pro is built for, and most of Moonshot’s headline benchmarks are its own. Artificial Analysis puts it at 55 on its Intelligence Index, behind Claude Opus 5.5 at 58.

How to use Kimi K3.1 Pro in a WordPress chatbot

MxChat reaches K3.1 Pro through OpenRouter:

1. Create an OpenRouter API key and add it under MxChat → Settings → API Keys.
2. In the chatbot’s model setting, choose moonshotai/kimi-k3.1-pro from the OpenRouter model list. The list is read live from OpenRouter, so new models appear without a plugin update.
3. Ask your bot the five questions your visitors ask most, check the answers against your knowledge base, and time them before you switch it on for everyone.

MxChat does not send a reasoning setting on the OpenRouter route today, so K3.1 Pro runs at its default, the slowest and most expensive of the four standard settings we measured. For a chatbot, that makes Claude Sonnet 5.5 (3.75 seconds) or a cheaper model the better choice for now. OpenRouter passes Moonshot’s price through without a markup on the model; our OpenRouter pricing test covers its fees against going direct. The MxChat documentation covers knowledge-base setup, and MxChat Pro adds the WooCommerce and live-agent features our test questions asked about.

Kimi K3.1 Pro against the other models we have tested

• Claude Sonnet 5.5 has the same list price, a 26% higher bill per answer on our prompt, and answers four times faster.
• Kimi K3 costs 35% more on Moonshot but a third as much on Relace, and answers in under five seconds on third-party hosts.
• GPT-6.1 Terra at low effort costs a third as much ($4.30) and answers in 1.44 seconds.
• Mercury 3 is the cheapest and fastest model we have tested, at $0.48 and 0.76 seconds per answer.

FAQ

How much does Kimi K3.1 Pro cost?

$2.00 per million input tokens, $0.20 per million cached input tokens and $10.00 per million output tokens on Moonshot’s API and OpenRouter. The Highspeed tier is $4.00 / $20.00. In our chatbot test that came to $12.76 per 1,000 answers with no cache, and $4.90 with Moonshot’s automatic cache.

Is Kimi K3.1 Pro cheaper than Kimi K3?

On Moonshot, yes: K3.1 Pro’s list price is a third lower ($2 / $10 against $3 / $15), and our per-answer cost fell 35%, from $19.54 to $12.76. But Kimi K3 is also sold by about 30 third-party hosts, and on the cheapest of them (Relace) it cost $4.71 per 1,000 answers.

Is Kimi K3.1 Pro cheaper than Claude Sonnet 5.5?

Per answer, yes, by about a quarter, even though the per-token prices are identical. Kimi’s tokenizer counted our 4,100-token prompt as 39% fewer tokens than Claude’s. Sonnet 5.5 answered more than four times faster.

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

Through OpenRouter, yes. reasoning: {"enabled": false} returned zero reasoning tokens on all eight of our calls, cut the cost by 30% and the median wait from 16.2 to 6.8 seconds, with every answer still correct.

Is Kimi K3.1 Pro Highspeed worth it?

Not for a chatbot. It produced tokens 2.4 times faster but reasoned almost twice as much, so answers arrived only 16% sooner while costing 2.5 times as much.

When are the Kimi K3.1 Pro open weights released?

Moonshot has said 4 November 2026. Once they are out, expect third-party hosts to list it the way about 30 already list Kimi K3, which has historically meant lower prices and faster answers than Moonshot’s own endpoint. We will update this page when that happens.

Does Kimi K3.1 Pro work with MxChat?

Yes, through OpenRouter. Add an OpenRouter key and pick moonshotai/kimi-k3.1-pro from the model list.

Method note: 96 chat-completion calls through OpenRouter from mxchat.ai’s WordPress server on 29 October 2026, 12 setups × 4 questions × 2 runs, all HTTP 200, providers pinned with fallbacks disabled where a provider is named. Costs are measured token counts at each route’s list price with no cache; repeat-pass costs are what OpenRouter billed. Times are the median wall-clock time to a complete, non-streamed answer. Prices read from OpenRouter’s model API on 29 October 2026; Moonshot’s launch pricing and the 4 November weights date as reported at launch.