Key takeaways
- AWS published its Bedrock access guide on November 10, 2026. The model card lists November 4 as the model launch date; this is not evidence of a newly trained model launched on November 10.
- Bedrock exposes Kimi K3.1 Pro Highspeed through global.moonshotai.kimi-k3.1, us.moonshotai.kimi-k3.1 and in.moonshotai.kimi-k3.1 cross-Region profiles, with a 1M-token context window and text/image input.
- Global Standard rates are $4 input and $20 output per million tokens; US and India geographic routes cost $4.40 and $22. Cache reads and writes have separate rates.
- AWS recommends Chat Completions or Responses over Converse. The model card documents multi-turn reasoning-block failures and rejects attached PDF/HTML document inputs on Converse.
- The blog’s always-on reasoning description conflicts with the live card, which documents none as an off switch. Validate the exact API and effort setting rather than assuming the routes behave identically.
What AWS announced—and when
AWS’s November 10, 2026 post says Moonshot AI’s Kimi K3.1 Pro Highspeed is available on Amazon Bedrock and explains its API, caching and service-tier options. The live model card lists November 4 as the model launch date. Our discovery radar carried a November 11 feed date, so we use the article’s displayed November 10 date for the guide—not a claim that availability began that day.
The practical news is another managed inference route for the million-token, vision-capable Kimi model. AWS lists tool calling, structured output, configurable reasoning and streaming. This is separate from Moonshot’s own API and the K3.1 Pro open-weight release we previously covered; it does not establish a new Highspeed weight download or a new benchmark result.
AWS attributes a 2–3× output-speed advantage over Kimi K3.1 Pro to Moonshot. That is a vendor-reported comparison, not an AccessAllGPT latency measurement or a guarantee of faster end-to-end agent completion. Reasoning length, tool time, caching and service tier can all affect the user’s wait.
The endpoint and model IDs builders can use
Use the bedrock-runtime endpoint. AWS recommends its OpenAI-compatible Chat Completions API; Responses is also supported, alongside Converse and Invoke. The OpenAI-compatible base URL follows https://bedrock-runtime.{region}.amazonaws.com/openai/v1. An OpenAI-compatible client still needs AWS-authorized credentials and the permissions for the chosen Bedrock profile; an OpenAI account key is not a substitute.
The global profile is global.moonshotai.kimi-k3.1. The geographic alternatives are us.moonshotai.kimi-k3.1 and in.moonshotai.kimi-k3.1. These are Bedrock identifiers, not Moonshot’s direct-API kimi-k3.1-pro-highspeed ID. AWS says existing Kimi K3 integrations can use the same endpoint and APIs with a model-ID change, but the documented limitations still need regression testing.
Global routing can use supported commercial AWS Regions. The US profile routes within the US geography; the India profile routes within Mumbai and Hyderabad. Choosing an entry endpoint in one Region is not the same as pinning inference to that single Region. AWS recommends GetInferenceProfile from the source Region to inspect current geographic destinations. Confirm account access and quotas before promising capacity.
Pricing depends on routing, tier and cache reuse
AWS’s pricing page and model card agree on Standard rates per million tokens: Global CRIS costs $4 input and $20 output; US and India CRIS each cost $4.40 input and $22 output. Global cache reads cost $0.20 and 30-minute cache writes $5; the geographic routes list $0.22 reads and $5.50 writes. These are Bedrock prices, not a universal Kimi price across providers.
AWS lists Flex at half the Standard rate, with potentially longer response times, and Priority at a 75% premium. The blog says on-demand quota is shared across Standard, Priority and Flex. A cheaper tier does not create separate capacity or establish a latency commitment. Record the resolved tier returned by the service when comparing runs.
Implicit caching is supported. Explicit caching requires at least 256 stable tokens per checkpoint, retains content for at least 30 minutes, and works with Chat Completions, Responses and Invoke—not Converse. Cache writes cost more than ordinary input, so a one-off long prompt is not automatically cheaper because explicit caching exists. Compare repeated-prefix reuse, billed writes and actual reads on the selected route.
The model card also lists a 10× output-token quota burndown: one output token consumes ten tokens from quota. That is a throughput-accounting rule, not a claim that the published output price should be multiplied by ten.
Read the model card before copying the quickstart
The sources disagree on reasoning. The blog says Highspeed always reasons and offers low, high and max, defaulting to max. The live model card documents none, low, medium, high, xhigh and max; it says none turns reasoning off and minimal is unsupported. We have not called the API to resolve the difference. Use the card as the more detailed current reference, then validate the exact request and response before relying on an off switch.
The card says Converse accepts reasoning_effort in additionalModelRequestFields without validating it, including values rejected by other APIs. Use reasoning_effort on Chat Completions or reasoning.effort on Responses for an explicit effort test rather than treating a successful Converse request as proof the setting took effect.
More consequentially, AWS documents an InternalServerException when earlier reasoning blocks are passed back in a multi-turn Converse request. It says this affects default configurations in frameworks such as LangChain and Strands Agents; its workaround is to remove prior reasoning blocks. Converse also rejects attached document inputs such as PDF and HTML. AWS recommends the OpenAI-compatible APIs where possible.
Video inputs are not supported on this Bedrock model. For images, the card says the detail setting is honored only by Chat Completions; Responses always processes images at high detail. A 1M-token context window therefore does not mean every input type, framework or fidelity control works uniformly across APIs.
Why this changes the buyer’s shortlist
For an AWS-centered team, the meaningful change is access through an existing cloud identity, billing and regional-governance stack rather than operating a separate direct-provider integration or a very large self-hosted model. AWS says inference stays within its data boundary, is not shared with Moonshot, is not used for training, and has zero retention and zero operator access for inference. Those are AWS assurances, not controls we independently audited; review the applicable terms and your application’s own logging separately.
Our recommendation is a small, non-sensitive comparison on the exact routing profile and API you intend to deploy. Measure task acceptance, multi-turn tool reliability, billed tokens, cache behavior and wall-clock latency with explicit reasoning settings. Compare Standard with Flex only on jobs that can tolerate delay. Keep the current route until the replacement passes representative work.
AccessAllGPT has not hands-on tested this paid model route. The next step is an account-specific pilot, not a provider migration based on the Highspeed name or a vendor speed multiplier. The AWS guide makes Kimi easier to evaluate for Bedrock buyers, while the model card makes clear that an endpoint-compatible call is not a behavior-identical integration.
A Bedrock pilot record
Use these fields for a proposed trial; this is not a completed AccessAllGPT benchmark.
Entries stay in this browser tab and are not submitted to AccessAllGPT. Blank responses are copied as [Unresolved].
Record the global, US or India inference profile, entry Region, permitted destination Regions and account quota.
Prefer Chat Completions or Responses; explicitly test the reasoning value and multi-turn tool flow you will use.
Record Standard/Flex/Priority, resolved tier, ordinary input, cache writes, cache reads, output and actual billed cost.
Compare task quality and end-to-end latency against the existing route using non-sensitive representative work; retain rollback.
Primary sources
Browse the publication-wide evidence index →
- Introducing Kimi K3.1 Pro on Amazon BedrockAWS · Reviewed: Displayed November 10 publication date; launch overview; Chat Completions and Responses examples; cross-Region profiles; reasoning; Flex; caching; availability · Retrieved · Supports: AWS’s availability statement, Bedrock API and profile identifiers, vendor-attributed speed comparison, service-tier and cache guidance, and inference data-handling assurances. Its narrower reasoning description differs from the current model card.
- Kimi K3.1 Pro - Highspeed model cardAWS Documentation · Reviewed: Model details and November 4 launch date; context; explicit caching; reasoning effort; Standard pricing; programmatic access; regional routing; quotas; usage limitations · Retrieved · Supports: Current vendor documentation for the 1M-token context, profile IDs, per-route prices, tier multipliers, cache minimum and retention, expanded reasoning controls, 10× output quota burndown, Converse failures, unsupported video and image-detail behavior.
- Amazon Bedrock pricing — Moonshot AIAWS · Reviewed: Moonshot AI on-demand pricing; Kimi K3.1 Pro - Highspeed Global, US and IN CRIS table; cache rates; Priority and Flex footnotes · Retrieved · Supports: Corroborates Standard per-million-token input, output, cache-read and cache-write rates across global and geographic inference, plus Priority’s 75% premium and Flex’s 50% discount.
Limitations
First-party source briefing retrieved November 12, 2026, not an independent availability, privacy or performance audit. We made no inference call, tested no account access, SDK, framework, reasoning setting, cache hit, quota or bill, and did not reproduce Moonshot’s speed claim. The blog and model card disagree on reasoning controls and show different announcement/model-launch dates. Availability, routing, quotas and prices can change. No separate Highspeed weight release is established here.
Disclosures
AccessAllGPT received no payment, credits or private vendor briefing for this report. AWS and Moonshot did not review or endorse it. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with AWS or Moonshot AI. Publication-wide relationships are listed on the disclosures page.
