Kimi K3 on Amazon Bedrock: pricing, caching and three ways to call it
AWS made Moonshot's Kimi K3 generally available on Bedrock on 2026-09-18: 1M context, native vision, $3 input / $15 output per million tokens on the Global endpoint, cache read $0.30, cache write $3.75, under the id moonshotai.kimi-k3. This page puts AWS's wording next to the other two routes - Moonshot's own API and an aggregator - because the weights are the same and the differences are rate limits, cache billing and operations.
Updated 2026-09-20
One key across Claude, GPT and Chinese models, billed per token.
- Billed per token — live rates on /models
- Pay by card (Visa / Mastercard / AMEX), Apple Pay, Google Pay or crypto
- Self-service: top up and activate instantly
Four hard numbers from the AWS docs
input/output per million tokens
Bedrock model card pricing: Global CRIS $3.00 in / $15.00 out; US CRIS is 10% higher ($3.30/$16.50). This is AWS's table, not Moonshot's list price and not a QCode tariff.
cache read (per million tokens)
Same table, Global CRIS row: cache read $0.30, cache write (30 min) $3.75; the US CRIS row is $0.33 and $4.125. The card adds that an explicit cache checkpoint needs at least 1,024 tokens, is kept for at least 30 minutes, and is currently available only on the Responses and Chat Completions APIs.
context window
The card states Context window: 1M tokens, aimed at long-running work over large repos and documents. Its modality matrix has to be read by the checkmarks, not by the row labels: on the input side only Text and Image are checked, while Audio, Speech and Video are not; on the output side only Text is checked (Embedding, Image, Speech and Video are not) - on Bedrock it is text-out only, images go in but never come out.
minimum tokens per cache checkpoint
Shorter prefixes never become a checkpoint. Implicit (automatic) caching is on by default; AWS says configuring explicit caching can raise the hit rate, and therefore cut latency and cost.
What K3 is, and what landing on Bedrock changes
The AWS card calls K3 Moonshot's "most capable open-weight model": native vision plus a 1M context window, GA on Bedrock on 2026-09-18 as moonshotai.kimi-k3. What changes is not the weights but the access layer: AWS endpoints and quotas, IAM and audit logs, region and CRIS tiers, a 45-day EOL notice, and an AWS price list. Explicit prompt caching is the headline feature - AWS calls it the first open-weight model with explicit caching on Bedrock. Parameter counts circulating in the community (2.8T) and training detail belong to Moonshot's accounting, and this page does not put them in AWS's mouth.
Release cadence
2026-09-18: the AWS What's New post "Kimi K3 by Moonshot AI is now generally available on Amazon Bedrock" shipped with a model card giving prices, modalities and an endpoint recommendation (bedrock-runtime for new apps, Chat Completions for K3). Check date for this page, 2026-09-20: prices and cache parameters are transcribed from that card's version of that day; re-verify before quoting them long term.
Three dates
2026-09-18: AWS announces GA of Kimi K3 on Bedrock; the model card gives 1M context, Global and US prices and explicit-cache parameters, and its modality matrix checks text and image on input and text only on output.
2026-09-20: this page checks QCode's catalog: kimi-k3 is listed with priced calls in the last 30 days, so the aggregator route is comparable to the AWS one rather than hypothetical.
2026-09-10 reference: Cognition's SWE-2 post states it is post-trained from Kimi K3 (2.8T) and publishes K3's own coding scores - K3 has been the base model everyone is building on this fortnight.
Documented vs not yet claimable
Written in the AWS docs
As of the 2026-09-20 model card: GA date, model id, 1M context, the four price rows (input, output, cache read, cache write) including the 10% US uplift, the 1,024-token minimum and ≥30-minute retention for explicit cache, the Responses/Chat-Completions-only restriction, the recommended endpoint, and the at-least-45-day EOL notice - all from that model card.
Do not write these as fact
As of 2026-09-20, "First open-weight model with explicit prompt caching" is AWS's wording and scoped to Bedrock, not to the industry. Parameter count (2.8T), leaderboard rankings and "strongest open weights" are Moonshot's or third parties' claims. AWS prices move with region and CRIS tier, and the model card adds two service tiers above Standard: Priority is billed at 1.75 and Flex at 0.5 times the Standard per-token rate; enterprise terms, concurrency ceilings and compliance zones were not checked and are not inferred here.
Bedrock vs an aggregator entry point
Amazon Bedrock (AWS terms)
moonshotai.kimi-k3; Global $3/$15, US $3.30/$16.50, cache read $0.30 / write $3.75. You get IAM, audit logging, region choice and a 45-day EOL notice, but explicit caching checkpoints are your job, and the invoice lands on AWS.
Aggregator (QCode)
The same model family is listed on QCode as kimi-k3 (checked 2026-09-20, with priced calls in the last 30 days): an OpenAI-compatible call that changes only the model id, no per-region price split and no manual cache checkpoints; cache behaviour and billing follow our pricing page. Current prices and limits come from /pricing and the live catalog.
Choosing a route: three checks
First, where the bill and the compliance boundary sit: if RAG and agents already live in AWS and you need data residency or one invoice, Bedrock is the natural side; if you want to hop between models from one place, an aggregator is cheaper to operate. Second, how you use caching: on Bedrock you must shape long prefixes into ≥1,024-token checkpoints and reuse them within 30 minutes, and the payoff shows directly at $0.30; treat the cache as magic and both routes bill you at full input price. Third, modality and endpoint: on Bedrock K3 takes text and images in and returns text only - the card spells out "Video inputs are not supported". Multi-turn calls through Converse raise an InternalServerException when earlier reasoning blocks are included (the default LangChain / Strands setups hit this), which is why AWS prefers the OpenAI-compatible Responses / Chat Completions APIs; for mixed text-and-image requests the card suggests placing image blocks before text blocks.
The QCode situation
kimi-k3 is in QCode's model catalog (checked 2026-09-20) with real priced calls in the last 30 days, called the same way as other models by swapping the model id. This page deliberately does not quote a QCode unit price for K3 - price, cache billing and rate limits come from /pricing and the live catalog. To compare AWS with an aggregator, run the same prompt template on both, measure the cache hit rate, and then read the two invoices.
Questions
Is K3 on Bedrock the same model as Moonshot's own API?
Same weight family, different access layer: AWS publishes it as moonshotai.kimi-k3 with its own prices, caching semantics, limits and compliance terms, so behaviour and billing are not identical to Moonshot's hosted endpoint.
Does the 1M context cost extra?
The card lists 1M as the context window with no separate long-context surcharge, but every input token is billed unless served from cache (read $0.30, write $3.75) - how you reuse long prefixes sets the real cost.
What does explicit caching actually require?
Per the docs: at least 1,024 tokens per checkpoint, retained for at least 30 minutes, and only on the Responses and Chat Completions APIs. Put the stable prefix first and the variable part last.
What is K3 good at, and where is it weak?
Vendor and third-party tables put it on long-context coding and visual input; the weakness has to be measured on your tasks. As a reference, Cognition's 2026-09-10 table scores base K3 at 44.2% on FrontierCode 1.1 Main and 21.5% on Terminal-Bench 4, against Fable 5.1's 50.9% and 55.8% - a vendor table, not an independent ranking.
Could K3 be retired on Bedrock?
AWS's lifecycle text for the entry is "not sooner than: Not Applicable, at least 45 day EOL Notice will be provided" with no EOL date set. That is a notice period, not a promise of permanence.
Can I use both routes?
Yes, and it is a common split: AWS for residency and one invoice, an aggregator for side-by-side model switching and fallback. Configure limits, cache semantics and logging separately - do not assume one code path behaves identically on both.
Sources
AWS What's New, "Kimi K3 by Moonshot AI is now generally available on Amazon Bedrock", 2026-09-18, https://aws.amazon.com/about-aws/whats-new/2026/09/moonshot-ai-kimi-k3-on-amazon-bedrock/; Amazon Bedrock User Guide model card "Moonshot AI — Kimi K3", https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-moonshot-ai-kimi-k3.html (both fetched 2026-09-20). Coding-table figures cited for contrast come from Cognition, "Introducing SWE-2", 2026-09-10, https://cognition.com/blog/swe-2 (a vendor table). QCode catalog listing and 30-day priced usage are our own checks.
Compare on one template before you commit
kimi-k3 is callable on QCode (checked 2026-09-20, priced calls in the last 30 days); prices and limits come from /pricing and the live catalog.
Read next
Kimi K3 vs GPT-5.6 Sol
How the two models we do sell compare on coding and cost.
GLM-5.3 vs Kimi K3
Two open-weight flagships compared, including base model and limits.
QCode pricing guide
How an aggregator bills: official price times service rate, plus caching and call accounting.
Not affiliated with Amazon or Moonshot AI. AWS-side prices, cache parameters and lifecycle terms are transcribed from documentation fetched 2026-09-20; coding-benchmark numbers are vendor-reported. QCode does not provide Amazon Bedrock and does not quote its prices.