Cheap Kimi API: What the Price Actually Depends On

Key takeaways

• Kimi is one of the lower-cost model families you can call. The trap is that "cheap" is a per-token claim, while your bill is input plus output, with a cache discount on top — three numbers, not one.
• The official documentation is explicit about the shape: input and output are billed separately, repeated prefixes are cached, and on the K3 series cache writes are billed separately by time-to-live tier. A quote that mentions only one rate is not answering the question.
• Which variant you call matters as much as price: a code-tuned line and a general chat line have different pricing and behave differently on the same prompt.
• Kimi's platform documents both an OpenAI-compatible entry and an Anthropic-compatible Messages entry. On KuaiAPI the client split still matters: Claude Code is served by Claude-family groups, while Kimi is offered on the OpenAI-compatible side — check the group mapping instead of assuming a base URL swap is enough.
• Before you pay, ask for three things: the group's multiplier against the official rate, the cache line, and the policy on failed requests.

Why "cheap Kimi API" is a fair question

The Kimi line has become a default cheap option for teams that are tired of paying frontier prices for routine work. Searching for a cheap way to buy it is a reasonable instinct — and the market has answered: relays that speak the OpenAI shape resell Kimi, and several of them rank on the first page for exactly this phrase.

But price shopping without the pricing structure gets you a number you cannot check. The official documentation describes the structure plainly, and it has three parts.

The bill has three parts, not one

Input and output are billed separately. This is the single biggest source of surprise. A coding session that reads a large file and writes a short patch is mostly input; a summarisation job is mostly output. Two workloads with the same token count can land far apart.

Repeated prefixes are cached, and caching has a price of its own. Kimi's API caches repeated request prefixes in two time-to-live tiers, and for the K3 series the cache write is billed separately by tier. That means a prompt that repeats a long system block is cheaper on the second call than the first — and a prompt that keeps rewriting that block never gets the discount.

Which variant you call changes the price and the behaviour. The current Kimi line includes a flagship general model, a code-tuned variant, and an earlier general model. Calling the code-tuned one for prose, or the general one for a large refactor, is a way to pay more for a worse result.

Three questions that decide your bill

1. What is your output share? If your traffic is mostly generation, output pricing dominates; if it is mostly reading code or documents, input and cache behaviour dominate. Optimise the one that is actually large.
2. Does your prompt have a long stable prefix? If yes, cache hits are where the money is. If you rebuild the system block on every call, you are paying full input price every time and no discount will ever show up.
3. Are you calling the variant that fits the job? The cheapest token is the one you did not have to send because the right model answered on the first try.

The client question, where most people get it wrong

Kimi's own platform documents three compatible entries: OpenAI Chat Completions at https://api.moonshot.ai/v1, OpenAI Responses on the same base, and Anthropic Messages at https://api.moonshot.ai/anthropic. The question is therefore not "Kimi only speaks OpenAI"; it is "which protocol does your client speak, and which entry does your provider expose for this model".

On KuaiAPI the path splits by client. An OpenAI-compatible client — Cursor, Cline, Cherry Studio, Chatbox, most SDKs — points its provider at the OpenAI-compatible base URL and a Kimi model id. Claude Code talks the Anthropic protocol and is served by Claude-family groups. Swapping a base URL does not turn a Claude Code setup into a Kimi setup; check two things: the provider mode in the client, and whether the platform maps the model you want into the group behind that protocol.

How to check: in your client, look for the provider dropdown. If it says "OpenAI Compatible", you are in the right place for Kimi. If the only mode is an Anthropic one, that path is a Claude-family group — confirm with the platform before assuming Kimi is reachable through it.

How to compare two Kimi quotes without a spreadsheet

Ask for the quote in a form you can multiply. KuaiAPI publishes each group's multiplier against the official rate, and each model's input, output and cache-read rates separately — so a quote decomposes instead of being compared as a single number.

Then check two things that are not price at all:

• What happens to a failed request? On KuaiAPI, failed requests are not billed, and that is a published commitment rather than a favour.
• Where does the money go if you stop? The refund route returns to the original payment method, with a stated handling window. Read it before you top up, not after.

Checklist before you pay

#	Check	Good answer
1	Is the model list public?	Kimi ids readable without signing in
2	Is the multiplier published?	A number you can multiply against the official rate
3	Is cache priced separately?	Input, output and cache read each shown
4	Is the variant clear?	Flagship and code-tuned named, not blurred together
5	Is the protocol clear?	The provider mode is stated, so you know what will work

FAQ

Is the Kimi API free? No. There are free chat tiers on the consumer product, but the API is billed per token. Any page promising a free Kimi API key for paid endpoints is selling something else.

Is Kimi cheaper than DeepSeek? It depends on your input/output mix and on cache behaviour, not on the headline rate. Both families are cheap; the difference appears when you multiply each quoted rate by your own token shape. Cheap DeepSeek API works the same structure through on the other family.

Can I use Kimi in Claude Code? Kimi's own platform documents an Anthropic-compatible Messages entry, so the protocol is no longer the blocker it once was. On KuaiAPI, the Claude Code path is served by Claude-family groups and Kimi is offered on the OpenAI-compatible side; check the group mapping before pointing Claude Code at a Kimi model. If your only goal is to use Kimi, an OpenAI-compatible client is the shorter path.

Does Kimi work with my existing OpenAI SDK code? Yes — same SDK, same request shape, different base URL and model id. Nothing in your application code has to change.

Next step

The model plaza lists each group's multiplier and each model's input, output and cache-read rates, so a Kimi quote can be decomposed before you commit. The status page shows how each group is behaving at this moment, and the refund page is the written path referenced above.

For the neighbouring decisions, One API Key for All Models shows the client setup when one credential has to serve several tools, Cheap Claude API Key walks through the same checks on a premium family, and What Is an AI API Aggregator? covers the operator-level questions. The English documentation covers endpoints and key configuration.