Kimi K3 API Pricing: $3/$15, 1M Context, and the Real Open-Weights Decision

Decision first: Kimi K3 is a serious frontier-model candidate for long-horizon coding, knowledge work, multimodal analysis, and agent workflows. Moonshot prices the official API at $3 per million cache-miss input tokens and $15 per million output tokens; cache-hit input is $0.30. Artificial Analysis independently scores K3 at 57, measures roughly 62 output tokens per second, and confirms a 1M-token context window. That puts K3 close to the frontier while costing less than several premium APIs.

There is one important timing detail. K3 is available now through Kimi.com, Kimi Work, Kimi Code, and the Kimi API, but Moonshot says the full model weights will be released by July 27, 2026. As of this article's July 18 snapshot, it is accurate to call K3 an open-model release in progress—not a model whose full weights are already downloadable.

Kimi K3 pricing snapshot

Route	Input / 1M	Output / 1M	Evidence status
Kimi official API — cache miss	$3.00	$15.00	Official, checked July 18
Kimi official API — cache hit	$0.30	$15.00	Official; input cache only
ComputeUnion relay snapshot	Not yet safely verified	Not yet safely verified	No relay quote is promoted as confirmed

This separation matters. AIHubMix has published a K3 product route, but a product page alone is not enough to treat a relay quote as independently verified. ComputeUnion will add channel prices to the same detail page after the price collector can confirm the model ID, units, and returned route. Until then, the official price remains the only confirmed procurement baseline.

What does Kimi K3 cost per month?

The cache-miss formula is: input tokens ÷ 1M × $3, plus output tokens ÷ 1M × $15. The cache-hit scenario changes only the input rate to $0.30.

Monthly workload	Cache-miss cost	All input cache-hit	What to test
10M input + 2M output	$60	$33	Repository fit and answer acceptance
100M input + 20M output	$600	$330	Cache ratio and verbosity
1B input + 200M output	$6,000	$3,300	Concurrency, retries, and provider resilience

Output is the expensive side of the K3 bill. Artificial Analysis reports that K3 generated 130M output tokens during its Intelligence Index evaluation—well above the comparison median. A strong score can therefore coexist with a high successful-task bill if prompts invite overlong answers. Set output limits and measure accepted work, not price per token alone.

Kimi K3 vs GPT-5.6 Sol, Claude Fable 5, GLM-5.2, and K2.7 Code

Model	External Intelligence Index	Official input/output	Best decision role
Kimi K3	57	$3 / $15	Frontier coding, agents, knowledge work, 1M context
GPT-5.6 Sol	59	$5 / $30	Maximum general frontier capability
Claude Fable 5	60	$10 / $50	Premium long-horizon work and polished interaction
GLM-5.2	51	See current provider quotes	Lower-cost Chinese coding alternative
Kimi K2.7 Code	42	$0.95 / $4	Cheaper Moonshot coding workload

The scores above use the same Artificial Analysis methodology where available and remain separate from ComputeUnion price data. K3 is not the score leader, but its $3/$15 official rate creates a credible middle position: materially cheaper than Sol or Fable at the stated rates, substantially stronger than K2.7 Code on the dated external score, and more expensive than Moonshot's previous workhorse.

What K3 is actually built for

• Long-horizon coding: Moonshot explicitly targets repository-scale and multi-step engineering work.
• End-to-end knowledge work: K3 is positioned for research and document-heavy tasks that require planning and synthesis.
• Agent workflows: the launch evidence includes tool use, browser interaction, and long task horizons.
• Native multimodal input: K3 accepts text and images and keeps a 1M-token context window.
• Future self-hosting evaluation: the 2.8T-parameter release is strategically important, but infrastructure planning should wait for the actual weights and license artifacts.

Three reasons not to migrate blindly

1. Weights are promised, not yet available. Do not design a production self-hosting budget from an unreleased artifact.
2. Conversation state matters. Moonshot warns that preserved thinking history is important and that switching models mid-session can destabilize results.
3. K3 can be over-proactive. Strong system boundaries and tool permissions are necessary for agents.

Clear conclusion

Buy an API pilot now if you need frontier-class coding or knowledge work at a lower official token rate; wait before committing to self-hosting. K3's independent score of 57, 1M context, native vision, and $3/$15 official price make it more than a launch headline. The practical test is whether it beats K2.7 Code, GPT-5.6 Sol, or the current production model on successful-task cost after verbosity, retries, cache hits, and tool failures are counted.

Sources and method

Model architecture, positioning, availability, limitations, API ID, official price, and the July 27 weight-release plan come from the Moonshot Kimi K3 launch announcement and Kimi API platform. Intelligence, speed, latency, context, and benchmark-provider coverage come from Artificial Analysis. ComputeUnion keeps official, channel, and market evidence separate and does not invent a relay price when no route has been independently checked.