Moonshot AITextVision

kimi-k3

kimi-k3 is Moonshot AI's model at $3.00 input and $15.00 output per million tokens, with a 1000K context window. On a typical 3:1 input-to-output mix that blends to roughly $6.00 per million tokens, placing it in the premium band of the 79 models tracked here — cheaper than 13% of them.

Open cost calculator View provider Open official platform

MODEL OVERVIEW

Context window1000K

Input price$3 / 1M

Output price$15 / 1M

Output speed60 tok/s

Quality score95

Data updatedUpdated Aug 2026

Price verified2026-08-06 · Official

Price noteFlat across the full context; every request runs a thinking pass billed as output tokens

By price, kimi-k3 ranks 68th cheapest of 79 on input and 68th on output. Its quality score of 95 ranks 8th of 79, and on value (quality score ÷ output price) it ranks 67th of 79. Its blended rate sits 90% above the catalog average.

In real billing terms: a support chatbot handling 50,000 turns a month (2,000 input / 500 output tokens each) costs about $675 per month on kimi-k3, while a RAG workload of 20,000 queries (8,000 input / 800 output tokens each) runs about $720. Because output is priced 5.0× input here, long-generation tasks scale cost fastest — cap output length before you optimise anything else.

A 1000K context window holds roughly 1,333 pages of A4 text (at ~750 tokens per page), ranking 26th of 79 here. That is enough to drop in an entire technical manual or several source files at once, which suits long-document summarization and codebase-wide analysis. Output runs at about 60 tok/s (67th fastest of 79), acceptable interactively, though long responses will feel slow.

We track 2 Moonshot AI models, and kimi-k3's blended rate sits 56% above that provider's average. If your task tolerates a little less headroom, deepseek-reasoner-v4 (DeepSeek) blends about 91% cheaper at a quality score of 92, and is worth benchmarking first. Going the other way, gpt-5.6-sol scores 98 on quality for about 33% more.

The closest model in Moonshot AI's line-up is kimi-k2.6 ($0.95 input / $4.00 output, quality 90). Against it, kimi-k3 scores 5 points higher on quality, runs 25 tok/s slower (about 29%), has a 738K larger context window, costs 71% more blended. When two models sit this close on price, the deciding factor is usually the quality ceiling — run both against your real prompts before picking.

Our read on it

Quality 95 at $3/$15 — but the line that decides your bill is that every request runs a thinking pass first.

Read the billing model before the rate

kimi-k3 lists at $3/$15, flat across the entire context window. That flatness is itself worth noting: plenty of long-context models step up to a higher tier past some threshold, and this one does not.

It has a different characteristic instead. Every request runs a thinking pass, and those thinking tokens bill as output. Your real output volume is therefore systematically higher than the visible length of the reply. Multiplying $15 by your expected response length will underestimate the invoice — to forecast it properly you have to measure the thinking-token ratio on your own prompts first, then compute from that.

What a 95 puts it up against — and where it loses

Ninety-five points places it in the flagship band against gpt-5.6-sol (98, $4/$20), claude-opus-5 (97, $5/$25) and claude-fable-5 (96, $10/$50). kimi-k3 undercuts all three at $3/$15, and its 1,000K context and vision support give up nothing to them.

The rest of the comparison is less flattering, and leaving it out would be dishonest. gpt-5.6-terra also scores 95, costs $2/$12, carries a larger 1,049K context, streams at 110 tokens/second against kimi-k3's 60, and adds audio. On paper terra wins almost every column. OpenAI's o3, also 95, is cheaper still at $2/$8 — but with only 200K of context and 40 tokens/second.

Given terra exists, when is kimi-k3 still right

Not on rate. The case rests on three things a comparison table cannot show. Region: teams operating in Asia-Pacific with their own measurements of Moonshot endpoint latency and availability may see a very different picture from what US-hosted inference gives them. Vendor concentration: routing all flagship traffic through a single American provider is an operational risk that deserves to be named out loud rather than discovered during an outage.

And the billing structure itself. kimi-k3 is flat across the full context with no step up past a threshold, so the cost curve for long-context work stays linear and forecastable. Against that, if your cost model has no headroom at all, a model that does not force a thinking pass — claude-sonnet-5 at $2/$10 — produces a more predictable invoice. Measure first, then decide. This page can take you as far as knowing what to measure.

Estimated monthly cost

Monthly API spend across four common workloads, next to the cheapest model we track.

Workload	Usage assumption	This model	Cheapest model
Support chatbot	2,000 input / 500 output tokens per turn, 50,000 turns per month	$675	$3.00
RAG document Q&A	8,000 input / 800 output tokens per query, 20,000 queries per month	$720	$3.84
Coding assistant	12,000 input / 2,000 output tokens per request, 10,000 requests per month	$660	$3.20
Batch summarization	4,000 input / 400 output tokens per item, 200,000 items per month	$3,600	$19

Derived directly from the rates on this page. Excludes prompt caching, batch discounts and free tiers.

How it compares

The alternatives worth benchmarking alongside it, from the same dataset.

Model	Provider	Blended /1M	Quality	Price delta
kimi-k3	Moonshot AI	$6.00	95	—
deepseek-reasoner-v4Cheaper option	DeepSeek	$0.54	92	-91%
gpt-5.6-solHigher quality	OpenAI	$8.00	98	+33%
kimi-k2.6Same provider	Moonshot AI	$1.71	90	-71%
qwen3.5Best value overall	Alibaba	$0.11	91	-98%

Best for

ReasoningCodingLong Context

Modalities

TextVision

Where it fits — and where it doesn't

Good fit

Coding: writing, refactoring and debugging across multi-file project context.
Reasoning: multi-step decomposition, planning and tasks that need rigorous inference.
Long documents: a 1000K window ingests a full contract, paper or large codebase in one pass.
Multimodal input: handles images, screenshots and scanned documents directly.

Poor fit

High-frequency batch jobs: at $6.00 blended per million tokens, bills scale quickly at volume.

Frequently asked questions

How much does the kimi-k3 API cost?

kimi-k3 costs $3.00 per million input tokens and $15.00 per million output tokens. On a 3:1 input-to-output mix that blends to about $6.00 per million tokens.

What does kimi-k3 cost per month?

It depends on volume. Using the four worked scenarios on this page: a support chatbot runs about $675/month, RAG document Q&A about $720/month, a coding assistant about $660/month, and batch summarization about $3,600/month. Plug your own token counts into the cost calculator for a tailored figure.

How large is the kimi-k3 context window?

1000K tokens — roughly 1,333 pages of A4 text at ~750 tokens per page, ranking 26th of the 79 models tracked here.

kimi-k3 or deepseek-reasoner-v4 — which is better value?

deepseek-reasoner-v4 blends to about $0.54 per million tokens versus $6.00 for kimi-k3 — roughly 91% cheaper — at a quality score of 92 against 95. If that quality gap does not show up on your task, take the cheaper one; if it does, stay with kimi-k3. Benchmark both on your own data before deciding.

What is kimi-k3 best suited for?

Tagged use cases: reasoning, coding, long-context. Coding: writing, refactoring and debugging across multi-file project context. Less suitable: High-frequency batch jobs: at $6.00 blended per million tokens, bills scale quickly at volume.

How current is this kimi-k3 pricing?

Pricing here was last verified on 2026-08-06 against official documentation. llmprice.app runs a daily collection job, but providers can change rates between runs — confirm against Moonshot AI's official pricing page before you commit.