Short answer: OpenRouter listed Kimi K3 at $0.192 per million input tokens, $16.00 per million output tokens, and $0.182 per million cached-input tokens on November 13, 2026, with a 1,048,576-token context. In a workload of 1,800 input and 400 output tokens per request, output accounts for roughly 95% of the bill. Cap answer length and measure completion tokens before assuming Kimi is the cheap option.
| Route | Prompt | Completion | Source and meter |
|---|---|---|---|
| moonshotai/kimi-k31,048,576 context | $0.192 | $16.00 | OpenRouter model API, retrieved for this guide. |
| moonshotai/kimi-k3:batchbatch variant | $2.28 | $11.40 | The batch row lists higher input and lower output than the standard route. |
| moonshotai/kimi-k3.1-pro1,048,576 context | $2.00 | $10.00 | A pro variant at a very different price point. |
| kimi-api on firework-aiserverless | $3.00 | $15.00 | APIDir accepted observation checked 2026-08-22. |
| kimi-api on together-aiserverless | $3.00 | $15.00 | APIDir accepted observation checked 2026-08-22. |
An 83x input-to-output ratio
Kimi K3's listed rates are $0.192 per million input tokens and $16.00 per million output tokens. That is an 83x ratio, the second-largest in the six-route comparison on this site, behind GLM-5.2 at roughly 971x and far ahead of Qwen3.8 Max (3x) or GPT-5.4 (6x). Ratios matter more than absolute prices when your workload's token mix is lopsided, because they decide which meter controls the invoice.
For a document-analysis product that reads 20,000 tokens and returns 300, Kimi's input price is almost irrelevant: 20,000 tokens cost $0.0038 per request. For a chat product with 1,800 input and 400 output, output is 95% of the cost. The same model can be the cheapest and the most expensive option in a comparison table depending on which of those two products you are building.
The batch row is not a simple discount
The :batch variant lists $2.28 input and $11.40 output. Compared with the standard route, batch output is 29% cheaper while batch input is almost 12x more expensive. That is the opposite of the usual batch pattern and one of the reasons a family-level price summary is unsafe. If the variant is intended for bulk work with large prompts, its input rate can erase the output saving for document-heavy jobs.
Treat such a row as an observation to verify, not a rule. Retrieve the price again before use, check whether the variant identifier matches what your client sends, and reconcile a small invoice sample. A single misread variant can move a forecast by thousands of dollars at production volume, and the catalog is the authority, not a blog post — including this one.
Worked estimate: where the money goes
At 120,000 requests per month, 1,800 input tokens and 400 output tokens, with a 35% cache-hit rate: uncached input is 140.4M tokens at $0.192 ($26.96), cached input is 75.6M at $0.182 ($13.76), and output is 48M at $16.00 ($768.00). Total $808.72 per month. Input handling costs less than 5% of the bill.
Now cap output at 250 tokens instead of 400. Output falls to 30M tokens ($480.00) and the total drops to $520.72, a 36% reduction with no change of model or route. Reducing input by the same 37.5% would have saved about $15. Output caps are where the leverage is.
Cache reads barely move this bill
The cache-read rate of $0.182 is only 5% below the $0.192 input rate. A cache hit on Kimi saves almost nothing, which means prompt-cache engineering is low-value here. Teams that apply a generic 'cache everything' playbook from another provider will spend effort for a rounding error.
The right optimisation is different: shorten answers, avoid duplicate completion calls, and prevent agent loops from generating repeated output. Those controls attack the $16.00 meter directly. If a workflow genuinely needs long outputs, compare the total cost on a route with a cheaper completion rate before optimising input.
Instrument completion tokens per accepted answer before changing routes. A product that averages 400 output tokens and returns answers people accept is not the same buyer as one that needs 1,500 tokens to be useful, even when both call the same endpoint. The first can justify Kimi's input price; the second should compare completion rates first and treat the input bargain as a secondary benefit.
- Log completion tokens per accepted answer; that ratio is the whole budget.
- Set a hard maximum output length in the request and in the product copy.
- Deduplicate retries and agent steps that regenerate the same text.
- Treat long-document input as the primary use case, not as a free tier for verbose chat.
When Kimi is the right route
Kimi fits input-heavy jobs: long-document review, codebase understanding, retrieval-augmented analysis, and pipelines that produce short structured answers. The one-million-token context supports those jobs technically, and the input rate keeps them affordable. The maximum completion field of 943,718 tokens is a ceiling to respect operationally, not a target to design around.
It fits less well as a general chat backend where answers are long, because every extra paragraph is priced at the highest completion rate in this comparison set. Run a bounded sample of your own prompts, record completion tokens per accepted answer, and multiply. If the answer needs to be long, the route choice should follow the output rate, not the headline input bargain.
Test the output-token share yourself
Start with the retrieved Kimi K3 rates, then change the output cap to see how much of the bill it controls.
Planning result
Review the example inputs, replace each rate with the exact current value from the linked source, then calculate. APIDir does not send these values to a provider.
Questions developers ask
How much does Kimi K3 cost per million tokens?
OpenRouter listed $0.192 input and $16.00 output per million tokens on November 13, 2026, with a $0.182 cache-read rate and a 1,048,576-token context.
Why is my Kimi bill so high if input is cheap?
Output tokens cost 83x the input rate. A workload averaging 400 output tokens per request spends about 95% of its budget on completions.
Is prompt caching useful on Kimi?
Barely. The cache-read rate is only about 5% below the input rate, so a cache hit saves very little compared with providers whose read rate is 90% off.
Should I use the Kimi batch route?
Only after verifying it. The observed batch row lists higher input and lower output than the standard route, which is unusual and worth checking against your own invoice sample.
Sources and check dates
Every rate in this guide was retrieved from the linked page on Nov 13, 2026. Prices can change; recheck the source before purchasing.