cloudflare/ai workers-ai-provider@3.3.1

on GitHub

Patch Changes

• #599 c72bb8b Thanks @threepointone! - fix: route native unified-billing catalog slugs (e.g. deepseek/deepseek-v4-pro) through the run path

  workers-ai-provider@3.2.x began treating any "<vendor>/<model>" id as a third-party AI Gateway catalog slug. This regressed models like deepseek/deepseek-v4-pro that Cloudflare serves natively on the unified-billing run path (env.AI.run): without providers the provider threw at construction, and with providers: [openai] it routed to the gateway universal chat/completions (BYOK) endpoint and returned auth errors.

  deepseek/* is now on the unified run catalog (defaults to env.AI.run), matching pre-3.2 behavior. Unified-billing eligibility is decided per-model by Cloudflare's catalog, so BYOK-only deepseek models (e.g. deepseek/deepseek-chat) still work by opting into the gateway path per call (transport: "gateway" / byok). The rest of the OpenAI-wire long tail (mistral, perplexity, cerebras, openrouter, fireworks) is not on Cloudflare's unified run catalog — env.AI.run returns model-not-found for those — so they remain BYOK gateway-path providers.

  When no providers are configured, a bare "<vendor>/<model>" id is now routed through the same hardened run path as dynamic/* routes: the gateway defaults to the account's "default" gateway (third-party unified billing needs one), cacheTtl/skipCache/metadata/collectLog are applied to the gateway options, and delegate-only options (byok, transport: "gateway", fallback, resume, resume callbacks) now throw a clear error instead of being silently dropped.