Hi Vercel Support, I'm experiencing consistent HTTP 403 responses from AI Gateway before inference starts when using higher-capability OpenAI models. The same project, AI Gateway API key and runtime configuration successfully execute openai/gpt-5.4-mini, including normal token usage and billing. However: openai/gpt-5.4
- Azure upstream: HTTP 403 pre-inference
- Azure upstream retry: HTTP 403 pre-inference
- OpenAI upstream forced through providerOptions.gateway.only = ["openai"]: HTTP 403 pre-inference openai/gpt-5.6-terra
- OpenAI upstream forced through the same routing policy: HTTP 403 pre-inference In every failing request:
- inference never starts;
- no inference region is assigned;
- token usage is zero;
- cost is $0;
- failure occurs within approximately 0.3–0.5 seconds. The team is on Hobby and has available AI Gateway credits. The AI Gateway API key is active and successfully used by GPT-5.4-mini. Zero Data Retention is disabled, no model/provider allowlists or routing deny rules are configured, and BYOK is not enabled. Relevant Gateway generation IDs:
- gen_01M2CVFV5RYH7MQQ3TF0HT8YVP — GPT-5.4 / Azure / 403
- gen_01M2CX38QV02F7MPF32NGWVT8X — GPT-5.4 / Azure / 403
- gen_01M2D1KPA3PE54ZS5B3PY5V8SE — GPT-5.4 / OpenAI / 403
- gen_01M2D4SBF3YMKDXC7FKCYR25YZ — GPT-5.6-terra / OpenAI / 403 Could you check the server-side reason/code associated with these 403 responses and confirm whether there is any account, model, managed-credential, or provider authorization restriction affecting this team? According to the current AI Gateway documentation, Hobby teams can use available Gateway models with Vercel-managed credentials and BYOK is optional, so I would like to understand why these requests are being rejected before inference. Bgrs.