Kimi K3 arrives on Amazon Bedrock with 1M context and prompt caching

Moonshot AI's Kimi K3 is now available on Amazon's Bedrock, the AWS service that lets companies call AI models from their existing AWS accounts. AWS's model page describes K3 as Moonshot's most capable open-weight model, with image input and a context window of one million tokens. Moonshot's pitch is that teams get Bedrock's access, encryption and audit controls. Its standard price matches what other hosts have charged, and the page shows that how a developer calls the model decides which features work.

The standard price is $3 per million input tokens and $15 per million output tokens when AWS may route requests to any of its regions worldwide, the same floor SemiAnalysis found across K3's providers on OpenRouter in July. Keeping requests inside US or India regions for data-residency reasons costs 10% more. Repeated input read from cache costs $0.30 per million tokens, a tenth of the normal input rate. Caching happens automatically, but AWS recommends explicit caching, where the developer marks the reusable part of a prompt, because it raises the hit rate. A marked section must be at least 256 tokens and is kept for at least 30 minutes.

The catch is Bedrock's own Converse API, its generic interface for all models. With K3 it does not support explicit caching, it rejects attached PDF and HTML documents, and it fails on multi-turn requests that include the model's reasoning from earlier turns, which AWS says affects agent frameworks such as LangChain and Strands Agents in their default settings. AWS recommends the OpenAI-compatible Responses or Chat Completions interfaces instead. Two other defaults matter for cost and capacity: K3 reasons at its maximum effort unless told otherwise, which spends more time and tokens, and each output token counts as ten tokens against an account's usage quota. Video input is not supported on Bedrock.