Kimi K3.1 Pro on Amazon Bedrock: $4 in, $20 out per million tokens

Kimi K3.1 Pro landed on Amazon Bedrock this week with a dollar price attached, and the number is the interesting part: $4 per million input tokens and $20 per million output for the Highspeed version, with a half-price tier underneath it.

Kimi K3.1 Pro Highspeed is now served from Bedrock, at $4 in and $20 out per million tokens on the cheapest global inference profile, according to Bedrock's pricing page. The model is Moonshot AI's post-trained follow-up to Kimi K3 — same 3T-class open-weight architecture, same 1-million-token context, but trained with substantially more post-training compute aimed at long-horizon agentic work, fewer factual errors, financial modelling and code for game engines, 3D scenes and CAD scripts, per AWS's launch post. Highspeed is not a different model in any architectural sense; Moonshot describes it as the same weights served faster, at "2-3× the speed of K3.1 Pro" in its launch write-up.

The pitch to anyone already on Bedrock is one string. K3.1 Pro Highspeed uses the same endpoint, APIs and features as Kimi K3, so a team can switch by editing the model ID and keep its IAM policies, quota and logging intact. That is the least glamorous sentence in AWS's post and the most consequential one: it turns a Chinese open-weight model into a drop-in option inside an AWS contract, with data processed inside the AWS boundary, not shared with Moonshot and not used for training, zero data retention always on for inference and no operator access to prompts or completions.

The prices come in three tiers, and the spread is large enough to change how you architect around the model. Standard is the number above. Flex serves the same model at half the Standard rate — $2 in and $10 out — in exchange for potentially slower responses; AWS recommends it for evaluations, summarisation and background steps in agentic workflows, and you choose the tier per request, so one application can send latency-critical calls to Priority at a 75% premium ($7 in, $35 out) and bulk work to Flex. The catch is that on-demand quota is shared across all three tiers, so a Flex-heavy app eats the same allowance as a latency-sensitive one.

Regional profiles cost more for the privilege of staying put. The global profile is the cheap one at $4 and $20; the US-only and India-only profiles are $4.40 in and $22 out, with cache reads at $0.22 and cache writes at $5.50 per million tokens, against $0.20 and $5.00 globally. For a company with Indian data-residency obligations, that is a roughly 10% surcharge for keeping inference onshore — a mild premium by the standards of sovereign-cloud pricing, and a notable one given that AWS shipped an India-specific profile at launch rather than months later.

Two levers move the bill more than the headline rate does. The first is prompt caching: explicit caching lets a team mark where a stable prefix ends, so system instructions, tool definitions and repository conventions are not re-processed on every turn of an agent loop. AWS requires at least 256 tokens of stable content before a cache breakpoint, retains cached content for at least 30 minutes, bills cache writes above the standard input rate and cache reads at a discount, and — the detail that matters for capacity planning — cache reads do not count against the input tokens-per-minute quota. Explicit caching works on the Chat Completions, Responses and Invoke APIs but not on Converse, which is worth knowing before porting an old Converse integration.

The second lever is reasoning effort, and it is a billing trap for anyone migrating code. K3.1 Pro always reasons before answering, and the effort setting takes low, high or max — with max as the default when nothing is set. A team that ports a Kimi K3 integration without touching the request body will silently buy the most expensive reasoning budget on every call. AWS says a lower value is the most direct way to cut latency and output tokens on simple tasks, which is polite phrasing for: the default is the premium setting.

Compared with what else sits on the same shelf, K3.1 Pro Highspeed is priced as a premium model rather than a cheap Chinese one. Kimi K3 is $3 in and $15 out on the global profile; K3.1 Pro Highspeed is a third more on both sides for the same architecture plus a faster serving path. The older Kimi K2.5, still listed, runs $0.60 in and $3.00 out — roughly one-seventh the input price. Anyone who reads "open weights" as "cheap" should look at that ladder, because the Bedrock number reflects the model's capability tier, not its licence.

That is the honest framing of the open-weight argument. Downloadable weights put a ceiling on what any provider can charge, not a floor: self-hosting a 3T-class model is a capital expense, and most teams will not pay it. What open weights actually buy here is bargaining power at renewal time and an escape hatch when a vendor changes terms. Bedrock's $4/$20 is what convenience costs when you would rather not run the GPUs — and the same pattern is playing out across the Chinese open-weight tier, where Meituan priced LongCat-2.5 at 2 yuan in and 8 yuan out under an MIT licence a day earlier.

The counter-argument is that the speed claim has no independent measurement behind it. "2-3× the speed of K3.1 Pro" is Moonshot's own figure, repeated by AWS with attribution rather than verified by either party; no third-party throughput benchmark for the Highspeed SKU has been published, and output tokens per second is exactly the kind of number that moves with batch size, prompt length and load. The capability claims are better documented — Terminal-Bench 4.0 at 65.7% against GPT-6 Astra's 58.2% and Claude Opus 5.5's 66.4%, OfficeQA Pro v2 at 64.4, a claim-level hallucination rate that Moonshot says fell from 23.4% to 8.3% on its hardest internal set — but those are vendor-run numbers too, with harnesses and judges of Moonshot's choosing.

It is also worth remembering how fast the ground moved under this model. We wrote in October that Kimi K3.1 Pro scored 55 on the AI index, three points behind Opus 5.5; the harder suite that replaced that index puts the best open-weights model at 46, twelve points back. Serving the model on Bedrock does not change where it sits on capability. It changes who can buy it with a purchase order.

What to watch. Whether AWS lists the standard, non-Highspeed K3.1 Pro as a separate SKU — the launch post describes Highspeed as what is available, and a slower, cheaper sibling would be the more interesting product for batch work. Whether the $4/$20 holds once DeepSeek and Alibaba finish their next price move. And whether anyone outside Moonshot publishes a tokens-per-second figure for the Highspeed endpoint, because until then the 2-3× is a promise rather than a spec.

Does a 1-million-token agentic model at $20 per million output tokens clear your cost bar, or is self-hosting finally cheaper? Write to us via our contact page.