Chinese models in Claude Code and Codex

Moonshot AI · Coding and reasoning models

Using Kimi K3 with a coding agent

Moonshot's 2.8T-parameter model — the first open 3T-class release — with a 1M-token context and native vision, built for agent runs measured in hours rather than turns.

Not wired up

Not yet, and it's the priciest candidate here at $3 in / $15 out per million tokens, which puts it near frontier pricing rather than the usual Chinese-model discount. That can still pay off on long agent runs, where a cheaper model burns more tokens getting to the same place. It won't pay off on autocomplete.

The numbers

Built by
Moonshot AI
Released
Jul 16, 2026
Context window
1,048,576 tokens
Per 1M tokens (in / out)
$3.00 / $15.00
In HaryAI
Not wired up

List price from the vendor's own API — cache-miss input and output. Third-party hosts are often cheaper, and cached reads change the real number.

What it takes to use it

Weights went public on 26 July 2026, so there are two routes: Moonshot's own API, or a host serving the open weights. Both expose an OpenAI-compatible endpoint, which is all a coding agent needs. One note for anyone searching: this is Moonshot's K series. MiniMax's models are the M and H series, and the two get mixed up constantly.

Want this one?

Leave an email and we'll tell you if we wire it up. That's the whole message — there's no newsletter attached to it.

Figures checked 2026-08-06 against vendor pricing pages and launch coverage. Models re-price often — if something here is stale, tell us.

Using Kimi K3 with a coding agent · HaryAI