Kimi K3.1 gives Bedrock agents a million-token window—but the API choice changes caching and residency

AWS has added Moonshot AI’s Kimi K3.1 Pro Highspeed to Amazon Bedrock, giving builders another long-context model for agentic coding and document work. For teams building data agents, the notable number is a 1-million-token context window: that is enough room for unusually large schema catalogs, query examples, policies and execution traces. But the operational constraints matter more than the headline capacity.

Pick the API before designing the cache

The model supports Bedrock’s Responses, Chat Completions, Converse and Invoke interfaces. AWS recommends Chat Completions for this model, and its feature matrix reveals a practical reason to decide early: explicit prompt caching works with Responses, Chat Completions and InvokeModel, but not with Converse.

That distinction matters for schema-heavy agents that repeatedly send the same catalog descriptions, semantic definitions or policy prefix. AWS says explicit cache checkpoints require at least 256 tokens and remain available for at least 30 minutes. The published standard-tier prices are $4 per million input tokens globally, $20 per million output tokens, $5 per million cache-write tokens and $0.20 per million cache-read tokens. Reusing a stable prefix can therefore change both latency and cost, but only if the application uses a supported interface and keeps the reusable material stable.

Maximum reasoning is the default

Kimi K3.1 Pro Highspeed reasons by default. If an application does not specify an effort level, AWS says the model uses max; supported settings run from none through max. A production data agent should set that field deliberately rather than inherit the most expensive reasoning mode for every request. Schema lookup, SQL repair and final narrative generation may warrant different settings even inside one workflow.

The model supports structured outputs, client-side tool calling, invocation logs, guardrails and application inference profiles. It does not support Bedrock Knowledge Bases or intelligent prompt routing. Teams should not infer that adding the model automatically supplies retrieval over business metadata; grounding and tool orchestration remain application responsibilities.

Cross-Region is not in-Region

AWS’s launch says the model is available wherever Bedrock is available through cross-Region inference. The detailed model card is more precise: supported profiles are US geographic, India geographic and global cross-Region inference, while in-Region inference is not supported in the listed regions.

That is a deployment boundary, not a minor routing detail. A team with US or India residency requirements can choose the corresponding geographic profile. A team that requires processing inside one specific AWS Region cannot treat the regional runtime endpoint as proof of in-Region execution; Bedrock may route the request according to the selected inference profile.

Kimi K3.1 Pro Highspeed expands the model choices for long-context agents, but it does not erase architecture decisions. Before loading a large database catalog into that million-token window, choose the residency profile, API, cache layout and reasoning budget—and test SQL quality independently, because AWS’s launch does not publish a text-to-SQL evaluation.