Kimi K3.1 Pro Highspeed Lands on Amazon Bedrock: $4/$20 Global Pricing Matches Moonshot, Flex Tier Halves It

Background and the Pain Point

Kimi K3.1 Pro shipped open weights on November 4: roughly 2.8T total parameters and more than 1.5 TB on disk. For most companies, open weights do not mean usable. Self-hosting needs a multi-node GPU cluster, and compliance teams want data to stay inside their existing cloud boundary. By putting the Highspeed variant on Amazon Bedrock on November 6, AWS turns an open flagship into a managed API governed by IAM and audit logs.

Architecture Highlights

Per AWS and Moonshot, K3.1 Pro is a natively multimodal agentic model with a 1M-token context. The Bedrock version runs on cross-Region inference (CRIS) with Global, US and India routing profiles, under the model ID moonshotai.kimi-k3.1. The model card lists streaming, implicit and explicit prompt caching, client-side tool calling, structured outputs, configurable reasoning effort, Guardrails and invocation logs. Explicit caching works on the Responses, Chat Completions and InvokeModel APIs but not on Converse.

Benchmarks and Measured Results

AWS published no throughput or latency figures for Bedrock, so real Highspeed tokens per second are not officially disclosed. Model scores still come from Moonshot, such as 65.7% on Terminal-Bench 4.0 versus 66.4% for Opus 5.5, and are vendor-reported. Pricing is the real story: the Global tier at $4/$20 matches Moonshot's own Highspeed price, US and India routing cost 10% more, and Flex at half price comes to about $2/$10, the same as Moonshot's standard API.

Developer Quick Start

Teams already on AWS can enable the model in the Bedrock console and call global.moonshotai.kimi-k3.1; AWS recommends the Chat Completions API. For long-context coding agents, configure explicit caching (cache reads cost $0.20, about 5% of the input price). Set service_tier: flex on latency-tolerant batch jobs to save half, and pick US or India routing when data residency matters.