Kimi API pricing explained: what you actually pay for

How is Kimi API pricing structured?

Three numbers per model, all quoted per million tokens: cache-hit input, cache-miss input, and output. There is no subscription, no minimum and no free tier — the platform is pay-as-you-go, billed on tokens consumed.

Model	Input (hit)	Input (miss)	Output	Context
Kimi K3	$0.30	$3.00	$15.00	1,048,576
Kimi K2.7 Code	$0.19	$0.95	$4.00	262,144
Kimi K2.7 Code HighSpeed	$0.38	$1.90	$8.00	262,144
Kimi K2.6	$0.16	$0.95	$4.00	262,144
Kimi K2.5	$0.10	$0.60	$3.00	262,144

USD per million tokens, excluding tax. Source: Moonshot’s per-model pricing pages, verified 6 August 2026. Kimi K2.5 is closed to new accounts and sunsets on 31 August 2026.

Two structural facts worth absorbing:

• Output always costs more than input — between 4× and 5× the cache-miss input rate on every model. Verbose responses are expensive responses.
• K2.7 Code HighSpeed is exactly 2× K2.7 Code on every tier, for identical weights. You are buying latency, nothing else.

What is context caching and why does it dominate real bills?

Moonshot caches repeated prompt prefixes automatically. When a request begins with text the model has already processed, those tokens bill at the cache-hit rate rather than the cache-miss rate.

The gap is large. On Kimi K3 it is $0.30 against $3.00 — a factor of ten.

This matters more than model choice for a specific and common workload shape: agents. An agent sends the same system prompt, the same tool definitions and the same few-shot examples on every single call. That prefix is identical every time, so it caches. The variable part — the actual user turn — is usually small.

A worked example on Kimi K3, with 100,000 input tokens and 5,000 output tokens:

• Cold, 0% cache hits: $0.375 per request
• Warm, 80% cache hits: $0.159 per request

That is a 58% saving from nothing but prompt structure. No model change, no quality trade-off.

The practical instruction: put the stable part of your prompt first and keep it byte-identical between calls. A timestamp, a session ID or a randomised example ordering near the top of your prompt breaks the prefix match and quietly costs you the entire discount.

What do the three mistakes look like?

These are the ones that make a real invoice exceed a calculator’s estimate.

1. Leaving reasoning_effort on max with Kimi K3. K3 always reasons, and the depth control defaults to the most expensive setting. Reasoning tokens bill as output at $15.00 per million. On high-volume simple requests this can dominate your bill while producing no visible benefit. Set it explicitly to low unless the task genuinely needs depth.

2. Assuming the headline input price is what you pay. If your prompts are all different, you pay cache-miss rates on everything — the $3.00 figure, not the $0.30 one. Conversely, if you have structured your prompts well, budgeting at cache-miss rates overestimates by 2–3×. Model both ends.

3. Forgetting tax. Moonshot quotes every price excluding tax and calculates it at checkout based on your jurisdiction. This is not a rounding error in every jurisdiction.

Does file processing cost extra?

Not for the upload and extraction step. Moonshot documents the file-related interfaces — content extraction and file storage — as temporarily free. If you upload a document and extract its text, that call itself does not incur a charge.

The bill arrives at the next step. Once you pass extracted content into a model as input, it is billed as input tokens exactly like text you typed. A 200-page PDF is not free to reason over; it is free to extract.

Note the word “temporarily” in Moonshot’s own documentation. This is not a permanent guarantee, and it is worth re-checking before you build a product whose economics depend on it.

How should you actually estimate your costs?

Work from your own numbers, not from a per-token rate.

1. Estimate tokens per request. Roughly 3–4 characters per token for English prose; code and non-English text tokenize less efficiently. The token counter will estimate a real sample for you.
2. Split input from output. They are priced very differently, and most people underestimate output because reasoning tokens are invisible in the response.
3. Estimate your cache hit rate honestly. Varied one-off prompts: 0–20%. Agent with a fixed system prompt: 60–90%.
4. Multiply by realistic volume, then add tax.

The API cost calculator does steps 2 to 4 across every model at once, including the cache-hit slider — which is the input most calculators omit and the one that changes the answer most.

Is the Kimi API cheaper than the alternatives?

Often, but “cheaper per token” and “cheaper for your workload” are different questions, and we are not going to answer the first one with numbers we have not verified against each vendor’s current published pricing.

What we can say precisely, because it comes from Moonshot’s own documentation: within the Kimi lineup, the spread between the cheapest and most expensive model is roughly 4× on output tokens, and the spread between a cached and uncached input token on K3 is 10×. Both of those gaps are larger than the differences people usually agonise over when comparing vendors.

Get your own prompt structure right before you shop around. It is the cheapest optimisation available and it applies whichever model you end up using.