Back to index
OtherA few hours to integrate·Pay per token, pricing on the platform rather than the docs

Kimi Platform Documentation (Moonshot AI)

3.5

Competent, OpenAI-compatible documentation for a model family with genuinely long context and aggressive pricing. The docs cover the mechanics well and never mention what any of it costs.

What We Liked

  • Covers streaming, tool calling, JSON output and reasoning properly
  • OpenAI SDK compatibility means integration is close to trivial
  • Long context is documented as a capability, not just a headline number
  • Coding-specific model variants are distinguished clearly from general ones

What Could Be Better

  • No pricing anywhere in the documentation, which is a real gap
  • Domain moved from moonshot.ai to kimi.ai and old links persist
  • Little guidance on when long context beats retrieval
  • Data handling and residency questions are not addressed in the developer docs

Detailed review

Moonshot AI has become one of the more serious Chinese labs to pay attention to, and the Kimi model family has earned its reputation on two things: very long context and prices that make the American frontier labs look expensive. The developer documentation is where you go to find out whether the engineering matches the reputation, and mostly it does. The model lineup as documented centres on Kimi K3 as the flagship, a very large mixture-of-experts model with a million-token context window, positioned for coding and knowledge work. Below that sit a coding-specialised variant with a 256K window supporting text, image and video input including a high-speed option, and a general-purpose model at the same context length with multimodal input and reasoning capability.

The distinction between the coding variants and the general ones is drawn clearly, which matters more than it sounds, because picking the wrong one is the most common cause of disappointing results and most providers muddle the guidance. The documentation covers the things that actually determine whether an integration works: API key setup, model selection, streaming, multimodal input, tool calling, JSON-structured output and reasoning behaviour. Tool calling in particular is documented with enough specificity to build against, which is not universal. OpenAI SDK compatibility means that for most teams the integration is a base URL change and a model name, and that compatibility is by now table stakes but still worth having.

The million-token context deserves a note, because the number gets used as a marketing weapon and the practical questions are subtler. Long context is not free. Latency rises, cost rises with tokens, and models do not attend uniformly across a very long window, so the honest engineering question is when stuffing a million tokens beats retrieving the right ten thousand. The documentation states the capability and does not help you reason about that trade-off, which is the single most useful thing it could add.

The glaring omission is pricing. There is nothing in the documentation about what any of this costs, no rate table, no tier structure, nothing. You have to go to the platform separately to find out. Given that cost is one of the two main reasons anybody evaluates Kimi in the first place, leaving it entirely out of the developer documentation is a strange decision, and it means the docs cannot answer the question most readers arrive with.

There is also a practical navigation issue: the platform has moved from platform.moonshot.ai to platform.kimi.ai, with the old domain redirecting, and a great deal of writing about Kimi still points at the old location or at model names that have since been superseded. Verify anything you read in a blog post against the current docs. The thing genuinely absent, and I would want it before recommending this for anything sensitive, is any treatment of data handling in the developer documentation. Where inference happens, what is retained, what the residency position is.

Teams in regulated environments will need those answers and will not find them here. Three point five. Solid, accurate, OpenAI-compatible documentation for a capable and cheap model family, undermined by the complete absence of pricing and by silence on the data questions that any serious buyer will ask.

[ final ]

The verdict.

Worth reading if you are evaluating alternatives to the American labs on cost or context length. Budget an extra half hour to find the pricing, because the documentation will not tell you.