AI Gateway is not passing response_format (json_schema, strict and non-strict) through to Inception for inception/mercury-3. The model ignores the schema and returns its own keys.

Repro: POST https://ai-gateway.vercel.sh/v1/chat/completions with model inception/mercury-3. The schema is {title: string, category: enum[web, mobile, game, other]}, required both, additionalProperties: false, and the prompt is “Name and classify this chat: ‘landing page for my bakery’”.

Expected: {"title": "...", "category": "web"} Actual (gateway): {"name": "Bakery Landing Page", "classification": "Web Development"}. This happened in 0 of 3 runs matching the schema, in both strict and non-strict mode.

Control: The identical request sent directly to api.inceptionlabs.ai (model mercury-3) matches the schema in 3 of 3 runs in both modes. AI SDK v7 generateText + Output.object through gateway('inception/mercury-3') also fails 3 of 3 with “No object generated: response did not match schema.”

Latency is also about 2× higher through the gateway (around 1.4s vs 0.7s direct) for these short calls.

Tool calling via the gateway works fine for this model. Only response_format is affected.