Open-weight models, fine-tuned and abliterated, served on our own hardware. They are not the newest weights on the market. They are the ones we can run fast, cheap, and without a refusal layer in front of them.
| Tool calling | Yes — Anthropic tool_use and OpenAI tool_calls |
| Vision | Yes |
| Structured output | JSON object and JSON schema |
| Streaming | Server-sent events on both endpoints |
| System prompt | Any length; yours is used verbatim |
| Model | Context | Price |
|---|---|---|
| GLM 5.3 | 1M | $4 / month |
| DeepSeek V4 Pro | 1M | $5 / month |
| KIMI K3.1 Pro | 1M | $6 / month |
Long technical requests they finish instead of summarising. Code work where they have to run tools and keep going. Instructions taken literally rather than improved on your behalf.
They will not tell you that a request is problematic. They will not offer three safer alternatives. They are not reasoning models tuned for math olympiads, and they do not always know what shipped last month — the weights are not fresh. If you need the newest benchmark scores, you want a bigger provider.
We run one machine per shard and the weights are not new, so a flat
monthly price covers what it actually costs us. There is no frontier lab inside the
price — that is why it is a fifth of Claude Pro and 10× Claude Pro plan usage. The
trade is capacity: the pool is shared, and during peak hours trial keys wait behind
each other. That is what the 429 means — see status.