Kimi K3’s API looks cheap on paper, but hidden fees and missing technical details can make it far more expensive than the headline numbers suggest.

What the hype skips

Moonshot AI’s launch material boasts a 2.8 trillion-parameter model that is “cheap” and “open”. The press release also promises downloadable weights soon. Neither claim holds up.

  • Weights aren’t available – Moonshot set a July 27 2026 deadline for a public checkpoint. Until then users must call the hosted API, so the “open-source” promise delivers nothing usable.
  • 2.8 T parameters are not all active – The number describes the architecture’s total capacity. K3’s Mixture-of-Experts design routes each token through 16 of 896 expert sub-networks. Moonshot has not said how many parameters run for a request, making memory and compute estimates impossible.

The pricing that looks good – until you count words

Published rates:

  • Cache-hit input: $0.30 per 1 million tokens
  • Uncached input: $3.00 per 1 million tokens
  • Output: $15.00 per 1 million tokens

Those rates look modest, but they ignore token volume.

A recent test measured K3’s output at 130 million tokens, more than double the 63 million produced by other leading models on the same tasks. The model’s verbosity inflates the output bill directly—twice the words means twice the cost. For code generation, essays, or chat agents, the $15 per million-token rate can dominate the spend.

What this means for different users

Developers and researchers

If you test K3 for coding assistance or data-driven research, impose hard caps on token output. An A/B test that pits K3 against a baseline model with identical output limits will show whether the higher token usage stems from the model or from prompt design.

Budget-conscious teams

The cache-hit discount applies only when the same input repeats. Craft stable prompt prefixes that produce identical input strings to push more requests into the $0.30 tier; this works only for highly repetitive workloads (e.g., templated queries). Most varied inputs will fall back to the $3.00 uncached rate.

Companies considering self-hosting

Waiting for the checkpoint release is wise. Hosting the model removes per-token fees but adds undisclosed licensing and infrastructure costs. Until the weights are public, the hosted API remains the only realistic option.

Production environments

Don’t switch just because the headline price seems lower than rivals. Run a pilot that measures cost per completed task, including hidden overhead from longer responses, instead of cost per token. Only then can you decide if K3 saves money.

What to watch next

  • July 27 2026 checkpoint release – weights slated for release on that date.
  • Active-parameter disclosure – knowing how many of the 2.8 T parameters run per token would let users size hardware accurately.

Bottom line

K3’s advertised $15 per million-token output price tells only half the story. Its tendency to generate twice as many tokens as peers can wipe out any headline savings, especially for long-answer tasks. Until the weights drop and active-parameter counts are clear, treat K3 as a premium, potentially verbose service rather than a cheap alternative. Run token-budget tests, enforce output limits, and switch only after you’ve measured the true cost per finished job.