HAR-verified GPT-5.6 Thinking routing mismatch: Chrome resolves to GPT-5.5-mini while Edge stays on GPT-5.6 (Case 13542447)

Summary

I have a HAR-verified, browser/client-session-specific routing mismatch involving ChatGPT Business and GPT-5.6 Thinking/Sol.

This is not based only on model self-identification. The requested and resolved models were extracted from the successful POST /backend-api/f/conversation responses.

Issue chronology

This did not begin as one simultaneous fallback across every GPT-5.6 mode:

1. Earlier, I observed GPT-5.6 Pro being routed to GPT-5.5-mini.
2. Later, GPT-5.6 Thinking/Sol also began resolving to GPT-5.5-mini in Chrome. This later phase is independently verified by a sanitized HAR.
3. Approximately three minutes later, Edge InPrivate on the same account and workspace correctly resolved the same requested model to GPT-5.6 Thinking.

Because the Pro and Thinking/Sol fallback behaviors appeared at different times rather than together, they may involve different routing conditions or a backend state/rollout that changed over time. They should not be assumed to have one cause without checking internal routing logs.

Strict controlled comparison

Both verified requests used:

• the same ChatGPT Business account and workspace
• the same ChatGPT conversation
• the same endpoint: POST /backend-api/f/conversation
• the same requested/default model: gpt-5-6-thinking
• HTTP 200 responses
• requests approximately three minutes apart

Chrome — incorrect fallback

Time: 2026-08-21 08:41:34.659 (Asia/Shanghai)
requested model: gpt-5-6-thinking
default_model_slug: gpt-5-6-thinking
resolved_model_slug: gpt-5-5-mini
request_id: 37acecfc-d6d9-42b0-9c79-e6fa59d2e2d4

Edge InPrivate — correct routing

Time: 2026-08-21 08:44:45.154 (Asia/Shanghai)
requested model: gpt-5-6-thinking
default_model_slug: gpt-5-6-thinking
resolved_model_slug: gpt-5-6-thinking
request_id: 1b9b906f-da73-4e4f-afc5-106d95e55b88

There was no usage-limit, rate-limit, capacity, or fallback notice in Chrome.

Expected behavior

When the requested/default model is gpt-5-6-thinking, the backend should either serve that model or clearly disclose an intentional fallback and its reason. Equivalent requests from the same account and workspace should not silently resolve differently solely by browser/client session.

Actual behavior

Chrome silently resolved to gpt-5-5-mini, while Edge InPrivate correctly resolved to gpt-5-6-thinking about three minutes later.

Request

Please compare the internal routing logs for the two request IDs and determine whether this is:

1. a browser/client-session rollout or experiment,
2. an undocumented fallback condition,
3. a false-positive classification, or
4. a routing bug.

OpenAI Support Case: 13542447.

Both sanitized HAR files and the full comparison report were sent privately to Support. I am not publishing raw HAR data or account identifiers.

Possibly related in pattern to #34677, but this report is distinct because it contains a same-account Chrome-versus-Edge HAR comparison and because the Thinking/Sol fallback appeared after an earlier Pro fallback rather than simultaneously.

Start by comparing the internal routing logs for request IDs 37acecfc-d6d9-42b0-9c79-e6fa59d2e2d4 and 1b9b906f-da73-4e4f-afc5-106d95e55b88. Determine whether the differing resolutions reflect a client rollout, an undocumented fallback, a false positive, or a routing bug; done means identifying the cause and documenting the expected behavior.