Potential silent model rerouting: GPT-5.6 Pro behaves as Instant / GPT-5.5 Mini

Potential duplicates detected. Please review them and close your issue if it is a duplicate.

• #34676
• #33502

Powered by Codex Action

i also encounter this. when asked what model are you, the response is im 5.5-mini

Me too. Wonder when it can be fixed...or whether they're going to fix it

Thank God someone brought this up.

I feel like this has been happening for almost a month now — the Thinking mode is completely gone from my GPT chat window.

Yes, I’m on the free plan, so I know there are still limits. But I really miss how useful it used to be.

This is what my interface looks like. You can see that the Thinking mode option is simply not there anymore.

<img width="1067" height="518" alt="Image" src="https://github.com/user-attachments/assets/95beeb9a-1db5-402c-a1ab-6a1033cd1af5" />

I’m starting to believe that this issue may be related to the IP address being flagged. It most likely occurs among users in China who are using VPNs.

I’m starting to believe that this issue may be related to the IP address being flagged. It most likely occurs among users in China who are using VPNs.我开始怀疑这个问题可能与 IP 地址被标记有关。这种情况最有可能发生在使用 VPN 的中国用户身上。

Is that so, bro? If that’s the case, then I guess there’s nothing we can do about it.

That said, there is a workaround: press F12, enable the mobile view in the developer tools, and you’ll be able to select Thinking mode again.
Once you do that, even after switching back to the desktop interface, the blue “Think” tag will stay there and can still be copied.

I reproduced a superficially similar symptom on ChatGPT Web in Safari (macOS, ChatGPT Pro, 2026-07-29), but a controlled check did not establish silent rerouting:

• When I first inspected the affected conversation, its picker was on Instant, not Pro.
• After explicitly selecting Pro, the UI displayed Pro and Pro thinking throughout the response and showed no fallback or rate-limit notice.
• Earlier in the same conversation, when asked for its identity, the assistant had confidently claimed GPT-5.5-mini. It later retracted that claim and acknowledged that it could not reliably read the client-selected model.
• The assistant also invented an unsupported “internal route test” explanation for an arbitrary marker supplied by the user.

This is evidence of a reproducible model self-identification / unsupported routing-metadata claim issue, not proof of a backend model downgrade. It matches the distinction described in #33838: maintainers should verify requested/served model from server-side telemetry rather than treating the assistant’s self-report as authoritative.

No conversation ID, account identifier, raw network log, token, or other private diagnostic data is included here.

Follow-up after correcting the test methodology: the earlier OK-style control was too simple and should not have been used to rule out a downgrade.

I repeated the comparison in fresh ChatGPT Web conversations using the same nontrivial digit-DP prompt and a locally verified answer:

• Instant: service turn duration about 15.7 s; UI showed “thought for 15s”; answer was correct.
• Pro: service turn duration about 12.8 s; UI showed “thought for 11s”; answer was also correct.

In this controlled sample, Pro was actually faster than Instant. Therefore response speed—even a very fast response—is not a reliable identifier of the served model. Simple identity prompts in the originally reported conversation completed in roughly 2–5 seconds, but those prompts do not require Pro-level reasoning and cannot distinguish the route.

The user's report of frequent Instant-like behavior after selecting Pro remains unresolved. The strongest actionable request is still for OpenAI to inspect server-side telemetry for requested model, served model, and any fallback/routing reason. The current picker state is mutable and does not establish which model served an earlier turn.

No private conversation identifiers or raw diagnostics are included in this public follow-up.

I have a related but distinct HAR-verified reproduction. Earlier I observed GPT-5.6 Pro falling back to GPT-5.5-mini; later, GPT-5.6 Thinking/Sol independently began falling back in Chrome.

The later phase has a strict same-account, same-workspace, same-conversation comparison:

Chrome, 2026-08-21 08:41:34.659 Asia/Shanghai
requested/default: gpt-5-6-thinking
resolved: gpt-5-5-mini
request_id: 37acecfc-d6d9-42b0-9c79-e6fa59d2e2d4

Edge InPrivate, approximately three minutes later
requested/default: gpt-5-6-thinking
resolved: gpt-5-6-thinking
request_id: 1b9b906f-da73-4e4f-afc5-106d95e55b88

No limit/capacity notice appeared. The sanitized HAR files were sent privately under Support Case 13542447 and are not being posted publicly.

Because Pro fallback and the later Thinking/Sol fallback appeared at different times, I do not think they should be assumed to have the same cause without comparing internal routing logs.

Is this issue going on for so long? Users from UK started expereinceing this in the last 4 days since 20.08.2026. USing GPT 5.6 Sol frontend but backend resolves to GPT 5.5 mini and even VPn cant help on it. Any chance to fix this as this is a disaster!