Moonshot AI’s Kimi K3 became generally available on Amazon Bedrock on 18 September 2026, according to AWS’s own announcement and the model card in the Bedrock documentation. AWS describes it as Moonshot’s most capable open-weight model and the first open model with 2.8 trillion parameters, with native image input and a 1-million-token context window.
The part worth pricing out is what it costs to run, because the same model carries two different rates depending on where the request is routed, and a third and fourth rate depending on how much of your prompt repeats.
What it costs
All figures below are per 1 million tokens on the Standard tier, taken from the Bedrock model card:
| Inference option | Input | Output | Cache read | Cache write (30 min) |
|---|---|---|---|---|
| Global cross-Region | $3.00 | $15.00 | $0.30 | $3.75 |
| US cross-Region | $3.30 | $16.50 | $0.33 | $4.125 |
Two multipliers sit on top of those base rates. Priority — the tier that returns the fastest responses — is billed at 1.75× the Standard rate, so a 75% premium. Flex, for workloads that can wait, is billed at 0.5×: half price. AWS states in the blog post that the global profile costs about 10% less than a geographic one, which is the gap between the two rows above.
The cache column is the one that changes the arithmetic for a coding agent. AWS states that on a cache hit the cached input tokens are billed at the discounted rate and do not count against the per-minute input-token quota — so repeated runs with a stable system prompt and tool definitions get both cheaper and less rate-limited.
The catch that is documented
Kimi K3 supports both implicit and explicit prompt caching, but explicit mode has terms worth reading before designing around it: each breakpoint must cover at least 1,024 tokens, the entry is retained for at least 30 minutes, and in explicit mode the tokens written into the cache are billed at a higher rate. A prompt that changes every turn pays the write premium without ever earning the read discount.
The model card also documents two limits AWS recommends designing around rather than discovering in production:
- Prefer the OpenAI-compatible Responses or Chat Completions APIs over Converse.
AWS says Converse has known limitations with this model, including an
InternalServerExceptionwhen reasoning content from earlier turns is included in a multi-turn request — which AWS notes affects frameworks such as LangChain and Strands Agents in their default configurations — plus rejection of attached documents such as PDF and HTML. - Video input is not supported. Image and text input are, and AWS recommends placing image blocks before text blocks for combined requests, though it cautions the behaviour is prompt-dependent.
Availability is through US-geography and global cross-Region inference profiles. From a
Chinese-reader standpoint the practical constraint is the account rather than the model:
Bedrock billing requires an AWS account with a payment method AWS accepts, and the
permissions listed in the setup notes are bedrock:InvokeModel,
bedrock:InvokeModelWithResponseStream and bedrock:CallWithBearerToken.
Why the cloud listing matters
A model card on Bedrock is not just a distribution event — 澎湃新闻 reported on 22 September that the long-rumoured revenue-share arrangement between Moonshot AI and overseas cloud providers has now landed, with the cloud platform listing the model and splitting revenue based on call volume, and that Kimi K3 has also been listed on Alibaba Cloud’s Bailian platform under a similar arrangement. Treat the commercial terms as reported by the outlet rather than as confirmed figures; neither AWS’s blog post nor the model card states them.
For anyone serving an open-weight model themselves rather than renting it per token, the other side of that decision is where it runs: providers such as Vultr bill GPU capacity by the hour, and InterServer prices a plain VPS by the month — the comparison worth making is your own token volume against the hourly rate, not the headline price of either.