Kimi K3 price is $0.30 per million cache-hit input tokens, $3.00 per million cache-miss input tokens, and $15.00 per million output tokens through the official Kimi API. Those three rates matter more than one blended headline because a coding agent can repeatedly send the same repository context while producing a much smaller stream of new output.
This guide uses the rates Moonshot published for Kimi K3 on July 16, 2026. Check the official Kimi K3 pricing page before setting a production budget because provider prices can change.
Kimi K3 API price
| Token type | Price per million tokens | What it means |
|---|---|---|
| Cache-hit input | $0.30 | Repeated context that Kimi can reuse from cache |
| Cache-miss input | $3.00 | New context that must be processed |
| Output | $15.00 | New text, code, reasoning, and tool instructions generated by the model |
The tenfold gap between cached and uncached input is important. A stable coding session that keeps the same repository context can cost much less than a workflow that constantly rebuilds its prompt. Moonshot says its official API reaches a cache-hit rate above 90 percent in coding workloads. Your actual result depends on how consistently your tool preserves the prompt prefix and cache key.
Real Kimi K3 cost examples
| Example task | Input and output | Mostly cached | No input cache |
|---|---|---|---|
| Focused code fix | 100K input and 10K output | About $0.18 | About $0.45 |
| Large repository session | 1M input and 100K output | About $1.80 | About $4.50 |
| Long agent run | 5M input and 500K output | About $9.00 | About $22.50 |
These are simple illustrations. They assume every input token is either cached or uncached and do not include retries, tool traffic, provider markups, or separate product subscriptions. A real session usually lands between the two input columns.
Why output can dominate the bill
Kimi K3 always reasons at max effort at launch. That can be useful for long coding and knowledge tasks, but output is also the most expensive token class. Asking for short progress reports, preventing repeated explanations, and stopping a failing loop can save more money than trimming a small prompt.
The one-million-token window is a ceiling, not a target. Loading an entire repository without retrieval or file selection can raise cost and make the model search through irrelevant context. Start with the files and evidence the task needs. Expand only when the model reaches a real dependency.
API billing and Kimi Code are different
The official Kimi API is usage based. Kimi Code membership plans package model access and quotas for the coding product. A developer calling kimi-k3 from a server should plan around API tokens. A person using the Kimi Code app should review the membership limits shown in that product.
How to control Kimi K3 spending
- Keep one stable session for related work so the prompt prefix can stay cacheable
- Use a persistent prompt cache key in agent workflows when the API client supports it
- Limit output to the code, decision, or report the next step actually needs
- Set task budgets and stop conditions before starting a long autonomous run
- Track cached tokens, uncached tokens, and output separately
- Test a representative repository before forecasting monthly spend
Price is only one side of the decision. Our live Kimi K3 benchmark page tracks quality and evidence, while the Kimi K3 and MiniMax M3 comparison looks more closely at value among open-model contenders.