Models
Every model we serve
3 models, one OpenAI-compatible endpoint. Pass the id as model;
nothing else in your code changes.
| Model | In / 1M | Out / 1M | Spends a plan | Pick it for |
|---|---|---|---|---|
| DeepSeek V4 Flash deepseek-ai/DeepSeek-V4-Flash-0731 | $0.084 | $0.168 | ×1 | The baseline. Cheapest we serve, free tier included, and plan sizes are quoted in it. |
| MiniMax M2.7 MiniMaxAI/MiniMax-M2.7 | $0.18 | $0.72 | ×2.9 | The depth option. Reach for it when DeepSeek is not holding up. |
"Spends a plan" is the burn rate shown in your console, measured against DeepSeek V4 Flash:
×2.9 means the same number of tokens uses nearly three times as much of your allowance.
zai-org/GLM-5.3-Flash is also served on the same
endpoint; check the console rate card for its current per-token price.
Prices verified October 13, 2026.
Plan sizes on /pricing.
The cheapest model we serve and the one plan sizes are quoted in — MIT-licensed, 304B parameters, within a couple of points of Opus 4.8 on agentic benchmarks.
The depth option of the catalogue — 10B active parameters and a SWE-Bench Pro score matching GPT-5.3-Codex.
Retired
Models no upstream serves any more. Their ids no longer resolve; the pages stay up as a reference for the prices and benchmarks they shipped with.
Retired from the router — no upstream serves it. Page kept for reference.
Retired from the router — superseded by GLM-5.3-Flash. Page kept for reference.
Which one should I use?
Start on DeepSeek V4 Flash. It is the cheapest, it is what every "up to N tokens" figure on the pricing page is quoted in, and on its own card it lands within a couple of points of Claude Opus 4.8 on the agentic benchmarks. Move a task up when the extra depth shows up in your evals: MiniMax-M2.7 spends your plan about three times as fast — so move tasks, not everything.
All of them answer the same endpoint, so switching is one string. If you are wiring this into an editor or a workflow tool, the setup guides have the exact fields per app.