kimi-k3 API — per-token pricing and code examples
Price per 1M tokens
Cache rates apply when the provider reports cache tokens.
79% offAbout this model
kimi-k3 is served through OpenAI Chat Completions, OpenAI Responses, Anthropic Messages. You pay $0.63 per 1M input tokens and $3.15 per 1M output tokens. That is 79% below the official price of $3.00 and $15.00. Cache reads cost $0.063 per 1M tokens when the provider reports cached tokens. It is included in budget packs from $9.99 up.
Where Kimi K3 fits
Moonshot AI designed Kimi K3 "for frontier intelligence scenarios such as long-horizon coding, knowledge work, and deep reasoning." Moonshot lists a 1M-token context window for kimi-k3. Output is controlled with max_completion_tokens, which Moonshot sets to 131,072 by default and allows up to 1,048,576. Kimi K3 understands images natively and Moonshot also lists video input, with text as the only output. One practical limit: Moonshot does not accept public image URLs for this model, so images go in as base64 data. Moonshot's Chinese guide now calls K3 its previous-generation flagship, and Kimi K3.1 Pro is described as a post-trained version of Kimi K3. The full kimi-k3 weights have been released.
Kimi K3 effort levels and fixed sampling
Kimi K3 always reasons, and Moonshot keeps its Preserved Thinking feature permanently on for this model. The thinking parameter used by the K2 family does not apply here; Moonshot's migration note says to remove that configuration. Depth is set with a top-level reasoning_effort that accepts low, high or max, and max is the default.
Sampling on kimi-k3 is fixed. Moonshot pins temperature at 1.0 and returns an error for any other value; top_p stays at 0.95, n at 1 and the penalties at 0. The safest approach is to leave all of these out of the request. Two more rules matter for agents. Switching the reasoning_effort level partway through a conversation invalidates prefix-cache hits, so pick a level per session and keep it. And the full assistant message, including reasoning_content and any tool_calls, must be sent back unchanged on the next turn.
What it costs on Cent AI
Kimi K3 is billed per token from your prepaid USD balance. Because the model always reasons and defaults to max, reasoning tokens, which Cent AI bills as output, can make up a large part of each request. Setting low for short, well-defined jobs is the most direct way to bring that down.
Moonshot caches context automatically for kimi-k3. The default cache lifetime is five minutes, and prompt_cache_options.ttl can extend it to one hour. On Cent AI, cache writes for both lifetimes are billed at the input rate shown in the price box on this page. Cent AI applies any cache rate only when Moonshot reports cached tokens in the usage of that request. Moonshot also notes that a partially matching block counts as a cache miss.
Using it from your tools
kimi-k3 is available on /v1/chat/completions, /v1/responses and /v1/messages. Cent AI does not change your parameters silently, which means a stray temperature: 0.2 from an old config is not dropped for you and comes back as an upstream_error with the model's status code. Tool calling supports tool_choice values auto, none and required. For JSON output, Moonshot supports response_format with json_schema and strict: true, and advises parsing only message.content. Here is a request in the /v1/chat/completions format:
{
"model": "kimi-k3",
"reasoning_effort": "low",
"prompt_cache_options": { "ttl": "1h" },
"messages": [{ "role": "user", "content": "Classify these tickets by area." }]
}
Use the one-hour lifetime only when the same long prefix will be reused across that window, such as a coding agent working through one repository for an afternoon.
Specification
- Context window
- —
- Input and output
- Text
- Endpoints
- /v1/chat/completions, /v1/responses, /v1/messages
- Smallest pack
- $9.99
Latency and error rate per model are on the status page. status.cent-ai.dev
What a request costs
Open the full calculatorCost per request
Estimate
kimi-k3An estimate at price version 3. Your bill comes from the usage the provider reports for each request.
Save $0.5642Code
curl https://api.cent-ai.dev/v1/chat/completions \
-H "Authorization: Bearer $CENT_AI_KEY" \
-H "Content-Type: application/json" \
-d '{"model": "kimi-k3", "messages": [{"role": "user", "content": "Hello"}]}'Set CENT_AI_KEY to a key from the dashboard. OpenAI SDKs use https://api.cent-ai.dev/v1; Anthropic SDKs use https://api.cent-ai.dev.
Price history
Version 3, in effect since Oct 31, 2026.
Effective
per 1M tokens
Inputold $0.63 new $0.63Cache readold $0.063 new $0.063Cache writeold $0.63 new $0.63Cache write 1hold $1.26 new $0.63Outputold $3.15 new $3.15Effective
per 1M tokens
Inputold $0.15 new $0.63Cache readold $0.015 new $0.063Cache writeold $0.15 new $0.63Cache write 1hold $0.30 new $1.26Outputold $0.75 new $3.15
Compare with
- hy4-previewInput$0.46704
- deepseek-v4-proInput$0.165
- gpt-6.1-solInput$0.58
- gpt-6.1-terraInput$0.15
Questions
How much does kimi-k3 cost on Cent AI?
$0.63 per 1M input tokens and $3.15 per 1M output tokens. Cache reads cost $0.063 and cache writes $0.63 per 1M tokens. Each request is rounded up once.
Is there a minimum charge?
A successful request costs at least $0.00002. A request that ends early, because the client left or the output was cut off, pays only for its tokens.
How do I pay?
Top up a prepaid USD balance with QRIS or USDT/USDC, or buy a budget pack once. Nothing renews automatically.
What if the provider does not report usage?
The request is billed from an estimate and marked estimated on the Requests page. An estimate never contains cache tokens.
Which endpoint do I call?
/v1/chat/completions, /v1/responses, /v1/messages. The OpenAI base URL is https://api.cent-ai.dev/v1.
Try kimi-k3 with one key
One prepaid balance for every model. Pay with QRIS or USDT/USDC.