Model availability changes. Shoud checks Groq’s model API instead of assuming a hard-coded model will always exist.
The backend fetches models available to the Groq account and caches the result for five minutes. It intersects that response with approved production GPT-OSS 20B and 120B chat models, plus Whisper speech models, the enabled Orpheus English preview voice model and Qwen 3.8 27B preview for image understanding. Enterprise and unreviewed models are excluded.
If Groq reports an approved model retired or unavailable, Shoud refreshes discovery and can retry once on an available approved sibling. When no approved model is available, the request fails clearly. A replacement with unknown pricing is not enabled automatically.
Fallback requests remain metered against your compute allowance. Adding a new model requires reviewing its production status, capabilities and cost, because an API model list alone does not establish those details.
