Kimi K3 on Amazon Bedrock: pricing, caching and access routes
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. The AWS price list fetched on 2026-10-19 lists Kimi 3 in a table of its own. The weights are the same, so what differs between Moonshot's own API, Bedrock and an aggregator is rate limits, cache billing and the operational terms.
Updated 2026-10-20
One key across Claude, GPT and Chinese models, billed per token.
- Billed per token — live rates on /models
- Pay by card (Visa / Mastercard), Apple Pay, Google Pay or crypto
- Self-service: top up and activate instantly
The hard numbers in the AWS docs
input/output per million tokens
In the Bedrock price list Kimi 3 has a table of its own: Global CRIS is $3.00 input and $15.00 output, US CRIS is $3.30 input and $16.50 output. This is AWS's table, not Moonshot's list price and not a QCode tariff.
cache read (per million tokens)
The cache rows of the same table: Global CRIS is cache read $0.30, cache write (30 min) $3.75; US CRIS is cache read $0.33 and cache write $4.125. The Kimi 2 table lists input and output only, with no cache prices.
context window
The model card states a context window of 1M tokens, aimed at long-running work over large repositories and long documents. The modality matrix has to be read by its checkmarks: only text and image are checked for input, while audio, speech and video are not; for output only text is checked — on Bedrock it is text-out only, images go in but never come out.
minimum tokens per cache checkpoint
This is the threshold from the AWS docs (checked 2026-10-19): a prefix shorter than it never becomes a cache checkpoint. The docs also say that implicit (automatic) caching is on by default, and that configuring explicit caching raises the hit rate, which lowers latency and cost. Explicit caching is currently available on the Responses, Chat Completions and InvokeModel APIs; Converse does not support it.
What K3 is, and what landing on Bedrock changes
K3 is Moonshot's open-weight flagship, and the AWS model card calls it 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. The parameter count circulating in the community (2.8T) and the training details are Moonshot's own account and have nothing to do with the AWS price list.
Official updates
2026-10-19: the price list now lists Moonshot AI with Kimi 3 in a table of its own - the Kimi 3 table lists the Standard tier only, and the price page has separate Flex and Priority service tiers whose discounts are noted in the page footnotes. The model card shipped with the AWS What's New post gives modalities and an endpoint recommendation (bedrock-runtime for new apps, Chat Completions for K3).
Timeline
2026-09-18: AWS announces GA of Kimi K3 on Bedrock; the model card gives 1M context, prices for the Global CRIS and US CRIS tiers and the explicit-cache parameters, and its modality matrix checks text and image on input and text only on output.
2026-10-19: checking QCode's catalog shows kimi-k3 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
From the Bedrock price list fetched on 2026-10-19: Kimi 3's price rows can be checked line by line - Global CRIS and US CRIS are priced separately, and the input, output, cache read and cache write prices per 1 million tokens are all listed; the page states that the figures shown are for the Standard tier, while the price page has separate Flex and Priority service tiers whose discounts are noted in the page footnotes. The model card adds the minimum checkpoint size for explicit caching, the ≥30-minute retention, the supported APIs, the recommended endpoint, and the at-least-45-day EOL notice.
Do not write these as fact
As of 2026-10-19, "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 price page and the model card add service tiers above Standard: Priority is billed at 1.75 and Flex at 0.5 times the Standard per-token rate, while the Kimi 3 table itself lists the Standard tier only. In the same Moonshot block, Kimi 2 has a table of its own that lists the regions; enterprise terms, concurrency ceilings and compliance zones are not published on the official pages.
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. On the price list of 2026-10-19, US CRIS is priced at cache read $0.33 and cache write $4.125. You get IAM, audit logging and region choice; explicit caching checkpoints are your job, and the billing lands on AWS.
Aggregator (QCode)
The same model family is listed on QCode as kimi-k3 (checked 2026-10-19, with priced calls in the last 30 days): an OpenAI-compatible call that changes only the model id, with no per-region price split; QCode bills per token, per-model rates are listed on /models, and rate limits come from /pricing and the live catalog.
Choosing a route: what to check
Start with where billing and the compliance boundary sit: if RAG and agents already run on AWS and you need data residency and a single billing account, Bedrock is the natural side; if you want to move between models from one place, an aggregator saves you operational work. Then look at caching: on Bedrock explicit caching wants long prefixes shaped into checkpoints of sufficient length and reused within a short window, and the payoff shows up directly in the cache read price. Finally the modality and the endpoint: on Bedrock K3 takes text and images in and returns text only, and the card states in as many words that video input is not supported; multi-turn calls through Converse raise an InternalServerException when earlier reasoning blocks are included (the default LangChain and Strands setups run into this), which is why the docs recommend the OpenAI-compatible Responses and Chat Completions; for mixed text-and-image requests the card suggests putting image blocks before text blocks.
The QCode situation
kimi-k3 is in QCode's model catalog (checked 2026-10-19) with real priced calls in the last 30 days, called the same way as other models by swapping the model id. QCode bills per token, per-model rates are listed on /models, and rate limits come from /pricing and the live catalog. Want to try a new model on QCode first? Pay-as-you-go Packs start at $4 for a one-time quota you use within 7 days (top-ups from $5).
Questions
Is K3 on Bedrock the same model as Moonshot's own API?
Same weight family, different access layer: AWS publishes it as an entry of its own, and its prices, caching semantics, limits and compliance terms all follow the AWS documentation, so behaviour and billing are not the same as on Moonshot's hosted endpoint.
Does the 1M context cost extra?
The model card states 1M as the context window of that entry, with no separate long-context surcharge; but every input token is billed unless it is served from cache and billed at the cache read or write price, so how you reuse long prefixes decides the real cost.
What does explicit caching actually require?
Per the docs: explicit caching is kept for at least 30 minutes, and the Responses, Chat Completions and InvokeModel APIs support it; there is no such option on Converse. Put the stable prefix first and the variable part last, and a hit is billed at the cache read price while a miss is billed at the input price.
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 writes a notice of at least 45 days for that entry, and no EOL date is set. That is a notice period, not a commitment to long-term availability.
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
Pricing and tier discounts: the Amazon Bedrock pricing page, fetched 2026-10-19, https://aws.amazon.com/bedrock/pricing/ ; context window, explicit-cache parameters and lifecycle terms: the Moonshot AI — Kimi K3 model card in the Bedrock User Guide, https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-moonshot-ai-kimi-k3.html ; the Kimi 3 launch: the AWS What's New post, https://aws.amazon.com/about-aws/whats-new/2026/09/moonshot-ai-kimi-k3-on-amazon-bedrock/ . The coding-benchmark figures used for contrast come from Cognition's official SWE-2 post (2026-09-10, a vendor table), https://cognition.com/blog/swe-2 . The QCode catalog listing and the 30-day priced usage are our own checks.
Compare on one template before you commit
kimi-k3 is callable on QCode (checked 2026-10-19, 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
GLM-5.3 and Kimi K3 side by side: two open-weight flagships, differing in base model, modalities and limits.
QCode pricing guide
How an aggregator bills: per token, with per-model rates listed on /models, plus how cache and calls are accounted for.
Crypto payment guide
Nine coins, 56 networks, zero platform fee for AI coding API credits.
Not affiliated with Amazon or Moonshot AI. The AWS-side prices, cache parameters and lifecycle terms quoted here are transcribed from public documentation fetched on 2026-10-19; coding-benchmark figures are vendor-reported. QCode does not provide Amazon Bedrock and does not quote its prices. The information follows the official pages; model availability and unit prices are listed on /models.