Chinese models in Claude Code and Codex
What actually works today, what doesn't, and what it costs.
We make tooling for Claude Code and Codex, so the question we get most often is whether you can point those tools at a Chinese model instead. This is the answer, model by model, with the setup work and the per-token price stated plainly. Where we have nothing to offer, it says so.
Coding and reasoning models
- DeepSeek V4 (Flash / Pro)Works today
The one model on this list you can use with Codex today. Flash runs $0.14 in / $0.28 out per million tokens — five to fifteen times less than OpenAI for the same work — and both tiers carry a 1M-token context.
- GLM-5.2Not wired up
Zhipu's 744B open-weights model scored 62.1 on SWE-bench Pro against GPT-5.5's 58.6, at roughly a sixth of the cost — the strongest case yet for running a coding agent on something other than a US frontier model.
- Kimi K3Not wired up
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.
Video models
These don't plug into a coding agent — different tool, different job. They're here because the same people keep asking about them.
- Seedance 2.5Not wired up
ByteDance's video model generates 30 seconds in a single pass at up to 4K, takes as many as 50 multimodal reference inputs, and can edit one region of a frame without regenerating the shot.
- MiniMax H3 (Hailuo 3.0)Not wired up
MiniMax shipped H3 the same day as Seedance 2.5 and took the opposite position: open, with the API live from day one. 2K video, native stereo audio, motion transfer and precise editing.
Figures checked 2026-08-06 against vendor pricing pages and launch coverage. Models re-price often — if something here is stale, tell us.