Kimi API, without the scavenger hunt
The model IDs, published K3 rates, a minimal request, and the operational details that affect cost.
Published K3 rates
Prices shown per one million tokens. Taxes and account-specific terms may vary.
| Model | Cached input | Cache-miss input | Output | Context |
|---|---|---|---|---|
Kimi K3kimi-k3 | $0.30 / MTok | $3.00 / MTok | $15.00 / MTok | 1,048,576 tokens |
Minimal Python request
OpenAI-compatible chat completion using the current K3 model ID.
from openai import OpenAI
client = OpenAI(
api_key="YOUR_MOONSHOT_API_KEY",
base_url="https://api.moonshot.ai/v1"
)
response = client.chat.completions.create(
model="kimi-k3",
messages=[
{"role": "user", "content": "Review this migration plan."}
],
extra_body={"reasoning_effort": "high"}
)
print(response.choices[0].message.content)K3 cached input is priced at one tenth of cache-miss input. Stable prefixes, clean session boundaries, and cache-hit telemetry are part of the application architecture.
Measure more than token price
For agentic workloads, compare total task cost, retry rate, tool-call success, wall-clock time, and acceptance-test pass rate. A lower per-token price can still lose if the agent loops or needs repeated recovery.
Reasoning effort
K3 supports low, high, and max reasoning effort. Treat effort as a workload setting: use lower effort for routine requests and reserve higher effort for tasks where deeper reasoning improves completion enough to justify the output cost and latency.
Migration warning
Do not replace only the model string and call the migration finished. Retest tool schemas, structured outputs, maximum token budgets, retry behavior, and any logic that assumed K2 thinking fields.