Three open-weight models, one key, one API. Switch by changing one string — or one click in your dashboard.
The workhorse. Fast, cheap, and it carries long agent chains without coming apart.
Use it as
"model": "glm-5.3" — see the docs.
Precise. It reasons more slowly and improvises less, but it is also the one that rarely breaks the code on step three.
Use it as
"model": "deepseek-v4-pro" — see the docs.
The longest and the most talkative. Holds a whole repository in context and stays coherent across long answers.
Use it as
"model": "kimi-k3.1-pro" — see the docs.
| Model | Price | Context | Good at |
|---|---|---|---|
GLM 5.3glm-5.3 | $4 / month | 1M | agentic work with tools, refactoring, scripts |
DeepSeek V4 Prodeepseek-v4-pro | $5 / month | 1M | debugging, math and formal tasks, surgical edits |
KIMI K3.1 Prokimi-k3.1-pro | $6 / month | 1M | whole repositories, long-form writing, voice and role |
Flat monthly price, nothing metered, and 10× Claude Pro plan usage — the comparison is on the pricing page.
1M is the window per request. A session that runs past what fits in one call is trimmed from the oldest turns first, so a session never dies — it just forgets the beginning. This is what every provider does at this size; we would rather say it up front than have you find it mid-task.
Tool calling on both API shapes, vision, structured output, streaming, and no refusal layer in front of any of them. The difference between them is temperament and window size, not whether they will answer you.
Get a key