AWS Puts Moonshot's Kimi K3.1 Pro Highspeed on Bedrock

Amazon Web Services made Moonshot AI's Kimi K3.1 Pro Highspeed available on Amazon Bedrock on November 10, 2026, pairing a 1-million-token context window with global, US and India inference profiles. Moonshot AI reports the Highspeed variant generates output two to three times faster than K3.1 Pro. The deployment lands six days after the underlying model's November 4, 2026 launch.

K3.1 Pro is a post-trained version of Kimi K3. It retains the 3T-class open-weight architecture and context window of its predecessor while using substantially more post-training compute, according to Moonshot AI. The company says the extra training targets long-horizon agentic work and fewer factual errors, with gains in financial modeling, office documents, and coding for game engines, 3D scenes and computer-aided design scripts.

Speed matters here because agentic coding chains many sequential model calls into a single task: reading files, planning a change, editing code, running tests, fixing failures. A higher output-token rate shortens the generation portion of each step.

What does migration require?

Existing Kimi K3 implementations on Bedrock switch to the new model by changing the model ID. The model supports tool calling, structured output, configurable reasoning, image input and streaming through the same endpoint and APIs. AWS demonstrates OpenCode, an open-source terminal coding agent, connected through Chat Completions and a Bedrock API key.

Developers can call the model through the OpenAI-compatible Chat Completions and Responses APIs, or Bedrock's Converse and Invoke APIs. AWS recommends Chat Completions.

The global cross-Region inference profile routes requests to any supported commercial AWS Region. The US and India profiles constrain routing to their respective geographies for data-residency requirements. India requests route between Mumbai and Hyderabad.

AWS says inference data stays within the AWS data boundary, is not shared with Moonshot AI and is not used to train the model. Zero data retention is always enabled for inference requests, and AWS operators cannot view prompts or completions during inference.

How much does it cost?

Standard-tier rates depend on the routing profile, priced per 1 million tokens:

• Global cross-Region: $4.00 input, $20.00 output, $0.20 cache read, $5.00 cache write (30 minutes)
• US cross-Region: $4.40 input, $22.00 output, $0.22 cache read, $5.50 cache write
• India cross-Region: $4.40 input, $22.00 output, $0.22 cache read, $5.50 cache write

The comparison with the outgoing model is unfavorable on headline rates. Bedrock's pricing page lists Kimi K3 at $3.00 per million input tokens and $15.00 per million output tokens for Global Standard inference. The new model charges more for both, though its global cache-read rate of $0.20 undercuts K3's $0.30.

AWS offers three service tiers to manage spend. Flex runs at half the Standard rate in exchange for potentially longer response times. Priority carries a 75% premium. The tier is selectable per request, so an application can use Standard for ordinary traffic, Priority for latency-critical calls and Flex for evaluations, summarization or background agent work.

Explicit prompt caching targets repeated context such as system instructions, tool definitions and repository conventions. Each breakpoint must follow at least 256 tokens of stable content. The feature works on Responses, Chat Completions and Invoke, but not Converse. Cache reads are discounted and don't count toward the input tokens-per-minute quota, while writes cost more than standard input.

Where are the limits?

The model card supports reasoning settings of none, low, medium, high, xhigh and max, with max used by default. Setting none turns reasoning off for a request. Lower effort can reduce latency and token consumption in exchange for task performance — a real consideration for cost-sensitive workloads given the default.

Converse carries additional constraints beyond its lack of explicit caching. It can fail with an internal server error when reasoning content from earlier turns is included in a multi-turn request, affecting frameworks such as LangChain and Strands Agents in their default configurations. AWS recommends removing prior reasoning blocks for multi-turn Converse requests. Converse also rejects attached PDF and HTML documents, and Bedrock does not support video inputs for this model.

There's a capacity wrinkle too. Each output token consumes 10 tokens from the applicable account quota on the bedrock-runtime endpoint, a separate constraint from the per-token bill.

For existing K3 users, switching requires little code change. The operating trade-off extends beyond compatibility: the potential gain is faster completion of sequential agent work; the counterweights are higher input and output prices, reasoning-driven token use and quota consumption. Flex, caching and effort settings are the controls AWS provides to balance those demands.