Kimi K3
Ships with product management skills built in — and open weights you can host yourself.
Released 16 Jul 2026; public weights published 26 Jul 2026. Newest entry here, so treat the numbers as fresher than the field experience.
- Maker
- Moonshot AI
- Context
- 1M tokens
- Pricing
- $3 / 1M tokens / $15 / 1M tokens
Cache hits around $0.30 / 1M input. Flat pricing across the full window, with no long-context surcharge — unusual, and it matters at scale.
Best for
- Standard PM artefacts fast — PRDs, OKRs, story maps, sprint plans
- Work that can't leave your infrastructure (weights are public)
- Long-context analysis where a per-token surcharge would bite
Prompting quirks
Start from a skill, not from a blank prompt
Kimi ships around 14 product management skills by default — PRD and spec writing, OKR coaching, user story mapping, weighted scoring, sprint planning, stakeholder comms. Pick the skill and give it one line. Writing an elaborate prompt from scratch here is doing work the toolset already did (see kimi.com/resources/product-manager-skills-for-agents).
Don't expect it to argue back
It challenges your input less than the top Claude and OpenAI models. Use it to produce the artefact, then take the artefact somewhere else to be stress-tested. Asking it "what am I missing?" is the one job it's worst at.
Thinking mode is always on
Like the other 2026 frontier models, reasoning isn't a toggle. Skip the "think step by step" scaffolding and state the goal, the constraints, and what done looks like.
No long-context cliff, so the whole window is usable
Where GPT-5.6 re-rates requests over 272K input tokens, K3 charges one flat rate across 1M. If your job is genuinely huge-context, price it here before you assume the US flagships are the cheap option.
Open weights change the question, not the prompt
Prompting it is unremarkable, which is the point. The decision is procurement, not prompt craft: this is the frontier-ish tier you can run where your data has to stay.
Copy-paste example
Tuned for Kimi K3. Swap in your own content where marked.
{paste the full set of transcripts, tickets or research docs}
Given everything above: extract the themes that appear in more than one
source. For each: the theme, how many sources mention it, and one
verbatim quote with its source. Ignore anything appearing only once.Facts verified against the official Moonshot AI docs on 19 August 2026.