Kimi K3.1 Pro on AWS (November 2026): Bedrock Adoption Guide

On November 6, 2026, AWS announced Kimi K3.1 Pro (Highspeed) on Amazon Bedrock. The model card dates the launch November 4, 2026. It is Moonshot’s current open-weight model on Bedrock: native vision, a 1M-token context, and the same security boundary as the proprietary models on the service. Kimi K3, launched September 18, 2026, stays Active and is the prior pin.

This is a routing note, not a new default agent model. Global $4 / $20 sits in the same published list-rate band as Claude Opus 5.5 ($4 / $20) and above the calculator rows for Claude Sonnet 5.5 ($2 / $10, Anthropic’s list until the Bedrock sheet confirms it), GPT-6.1 Sol US short context ($2.20 / $11), and Grok 4.7 Global ($2 / $6).

The same card rates, applied to one token mix:

for 1,000,000 uncached input tokens and 200,000 output tokens on K3.1 Pro Global Standard ($6 on Kimi K3).

$8

higher Global input and output than Kimi K3 at the Standard rate.

33%

cache read per million tokens on K3.1 Pro ($0.30 on Kimi K3).

$0.20

minimum for an explicit cache checkpoint on K3.1 Pro.

256 tokens

Price that mix on the Bedrock pricing calculator. The Kimi rows use Global CRIS Standard. The calculator Batch toggle is the site-wide 50% factor. The model card does not publish a Kimi batch rate.

What changed on November 6, 2026

AWS added Kimi K3.1 Pro (Highspeed) for agentic coding and document work, available through cross-Region inference in every Region where Bedrock is available. The card confirms explicit prompt caching, client-side tool calling, structured outputs, Guardrails, and invocation logs. Knowledge Bases, intelligent prompt routing, CountTokens, and watermarking are not supported. Video inputs are not supported.

The name on the card is Highspeed. The card does not publish an output-tokens-per-second figure, so this post does not quote one.

Specs that change the call

From the Kimi K3.1 Pro model card unless noted:

Opinionated take: use Kimi K3.1 Pro when the lane should stay on Moonshot’s current open-weight model inside Bedrock, with explicit prompt caching, for long-context coding or document work. Keep Sonnet 5.5, GPT-6.1 Sol, and Grok 4.7 where they already pass a bakeoff. Do not attach this model to a Knowledge Base. Call it with Chat Completions if the loop keeps reasoning blocks from earlier turns.

Trade-off you accept: a newer Moonshot pin and cheaper cache reads, against a one-third higher input and output list rate and a client change if the lane is on Converse today.

Prices: K3.1 Pro against the Kimi K3 pin

Standard tier, per 1M tokens, from the two model cards:

A cache-heavy prefix can narrow the gap, because the read rate went down while the uncached rates went up. It does not flip the $8 versus $6 example above, where most of the input is uncached and the output is large. Re-forecast from the card, then confirm the live Bedrock pricing page before you lock a budget. Run the Global rows in the Bedrock pricing calculator or the token cost calculator.

Invoke K3.1 Pro (Responses API, Global profile)

Context: Python with openai and aws-bedrock-token-generator, calling bedrock-runtime in us-east-1. The model ID is the Global profile, so the source Region is where you send the request and Bedrock routes it. This is the sample shape on the model card, with the Region set to us-east-1.

from aws_bedrock_token_generator import provide_token
from openai import OpenAI

region = "us-east-1"
client = OpenAI(
    api_key=provide_token(region=region),
    base_url=f"https://bedrock-runtime.{region}.amazonaws.com/openai/v1",
)

resp = client.responses.create(
    input="Summarize the rollback steps in this runbook.",
    model="global.moonshotai.kimi-k3.1",
)
print(resp.output_text)

Set reasoning on the API you actually call. The card names reasoning.effort for the Responses API, reasoning_effort for Chat Completions and InvokeModel, and says Converse will accept reasoning_effort inside additionalModelRequestFields without validating it. Leave the field unset and the card says the model uses max.

Profile IDs from the model card:

• Global: global.moonshotai.kimi-k3.1
• US geo: us.moonshotai.kimi-k3.1 (US-geography Regions only)
• India geo: in.moonshotai.kimi-k3.1 (Mumbai and Hyderabad)
• Base ID: moonshotai.kimi-k3.1

Prior pin: moonshotai.kimi-k3, global.moonshotai.kimi-k3, us.moonshotai.kimi-k3, in.moonshotai.kimi-k3.

Two further card limits that show up only after the first happy-path text call: video inputs are not supported, and mixed text-plus-image requests do better when the image block is placed before the text. The detail image parameter is honored on Chat Completions only. On the Responses API the card says images are processed at high detail.

Where Kimi K3.1 Pro fits

Pair agent lanes that move money or orders with AgentCore and a token budget. The Bedrock service page still shows a Converse tool loop because that is the AWS-native shape for models that support it. For this model, follow the card and prefer Chat Completions.

Rollout: one lane, then the next

1. Pick one Moonshot lane. A coding loop or a document pass. Leave Sonnet, GPT, and Grok IDs alone.
2. Freeze the prompts. Replay traces you already have. Do not tune the prompt while you compare K3 and K3.1.
3. Score three numbers. Quality pass rate, latency to a completed task, and cost per completed task. List price alone will lie if 3.1 spends fewer tokens, or if max reasoning spends more.
4. Move the client off Converse before the bakeoff, or the scores measure InternalServerException instead of the model.
5. Set a reasoning effort. Start at medium, or none when the request should not reason. Confirm output-token quota at the 10× burndown before you raise it to high or max.
6. Promote that lane only. Keep moonshotai.kimi-k3 in production until the scores land.

What to do this week

1. Copy the profile from the K3.1 Pro model card that matches your residency rule. Global is global.moonshotai.kimi-k3.1.
2. Smoke-test Chat Completions in a non-prod account with reasoning effort set to medium.
3. Grep callers for Converse, for reasoning blocks echoed into the next turn, and for PDF or HTML attachments on this model ID.
4. Price the lane on the Bedrock pricing calculator using the Global row, then apply US CRIS ($4.40 / $22) if the profile is us.moonshotai.kimi-k3.1. Ignore the Batch toggle.
5. Leave Kimi K3 pinned until the one-lane eval finishes. Talk to us about the bakeoff from the Bedrock consulting page.

What this post doesn’t cover

• A first-party FactualMinds latency or quality bakeoff of Kimi K3.1 Pro against Kimi K3, Sonnet 5.5, GPT-6.1 Sol, or Grok 4.7. We have not run that bakeoff. The $8 versus $6 figure is model-card arithmetic for one token mix.
• An output-tokens-per-second number. AWS calls the model Highspeed. The card does not publish a rate.
• A parameter count. The What’s New post attributes a size class to Moonshot. The model card does not state it.
• Audio or speech inputs and outputs. The card’s modality grid is a template, and the limitations section says video inputs are not supported. This post claims text and image only.
• Knowledge Base, prompt-routing, or batch pricing. The card lists the first two as unsupported and does not publish a batch rate.
• AgentCore packaging. Guardrails are supported on bedrock-runtime. That is not the same as a production agent loop.

Use the card rates and the Converse limits as the adoption frame, then validate on your own prompts and quota.

Related reading

• Third-party models on Amazon Bedrock Marketplace
• Claude Opus 5.5 on AWS (September 2026)
• AWS Bedrock Cost Optimization: Token Budgets and Model Selection
• Amazon Bedrock AgentCore: Production-Ready AI Agents
• Kimi K3.1 Pro model card
• Kimi K3 model card

Frequently asked questions