A new tier for coding and agentic work
v1.2.2 adds one tier and changes nothing about the existing ones.
onde-prism
onde-prism is a premium tier for coding and agentic work. It appears in /v1/models and takes the same requests as every other tier, streaming or not, with or without tools.
Two providers, same weights
The API of the company that built the model is tried first, and a third-party host of the same open weights follows as the fallback. Unlike onde-zenith, a request answered by the fallback comes from the same model.
Forced tool calls and token budgets
A request that sets tool_choice to anything other than "auto" skips the first provider and is served by the fallback. The model thinks before it answers and that thinking is billed as completion tokens, so leave room in max_tokens: a tight limit can run out before any visible answer.