I am experiencing a reproducible ChatGPT model routing issue on my Plus account.
When I manually select GPT-5.6 Thinking/High on desktop Chrome, the request metadata initially shows:
model_slug: gpt-5-6-thinking
default_model_slug: gpt-5-6-thinking
thinking_effort: extended
However, the final server_ste_metadata reports:
model_slug: gpt-5-5-mini
requested_model_experience: thinking
did_auto_switch_to_reasoning: false
is_autoswitcher_enabled: false
So this does not appear to be merely a UI or model-selector display issue.
The final server metadata reports the turn as gpt-5-5-mini, despite the earlier metadata requesting gpt-5-6-thinking.
I performed controlled A/B tests using:
- The same PC
- The same ChatGPT account
- The same Plus subscription
- The same network/IP
- The same prompt
- The same GPT-5.6 Thinking setting
- No web search or other tools
On Microsoft Edge, the exact same test correctly resolves to:
model_slug: gpt-5-6-thinking
requested_model_experience: thinking
The Edge User-Agent is clearly identified by:
Edg/152.0.0.0
On Google Chrome, the same prompt and settings resolve to:
model_slug: gpt-5-5-mini
even though the earlier turn metadata still reports:
model_slug: gpt-5-6-thinking
default_model_slug: gpt-5-6-thinking
thinking_effort: extended
The Chrome User-Agent is plain:
Chrome/152.0.0.0
without Edg.
I also ruled out normal Chrome profile corruption.
I launched Chrome using a completely separate temporary profile and disabled all extensions:
"C:\Program Files\Google\Chrome\Application\chrome.exe" --user-data-dir="%TEMP%\chatgpt-clean" --disable-extensions
After logging into the same ChatGPT account, the clean Chrome profile still resolved GPT-5.6 Thinking to:
model_slug: gpt-5-5-mini
with final server metadata containing:
user_agent: web_desktop
model_slug: gpt-5-5-mini
a32e6ebcb: 1
The most interesting result came from changing only Chrome’s User-Agent through DevTools.
I temporarily changed desktop Chrome’s User-Agent to an Android mobile User-Agent.
ChatGPT then classified the same browser as:
is_mobile: true
is_mobile_web: true
is_mobile_web_android: true
is_android: true
is_windows: false
user_client_type: android_web
Immediately after doing this, the final server metadata changed to:
user_agent: web_android
model_slug: gpt-5-6-thinking
a32e6ebcb: 8
Even more strangely, after restoring Chrome to its normal/default desktop User-Agent, the routing remained fixed.
The browser was again correctly classified as:
is_windows: true
is_mobile: false
is_android: false
user_client_type: desktop_web
but the final server metadata now remained:
user_agent: web_desktop
model_slug: gpt-5-6-thinking
a32e6ebcb: 8
So the sequence I observed was:
Desktop Chrome
→ a32e6ebcb: 1
→ gpt-5-5-mini
Change UA to Android
→ a32e6ebcb: 8
→ gpt-5-6-thinking
Restore normal desktop Chrome UA
→ a32e6ebcb remains 8
→ gpt-5-6-thinking
I do not know what the undocumented a32e6ebcb field represents, so I am not claiming that it is the cause of the routing issue.
I am only reporting the correlation visible in the server metadata.
This behavior makes a normal account-level quota or risk-control explanation seem unlikely, because the same account can receive GPT-5.6 on Edge at the same time that Chrome receives GPT-5.5-mini.
Changing only the reported client identity also caused GPT-5.6 routing to immediately return.
My current suspicion is that this may involve one of the following:
- Browser/client-specific rollout behavior
- A routing cohort or experiment
- Cached backend routing state
- A feature-flag interaction
- A browser/client-specific routing bug
However, only OpenAI can confirm the actual cause.
The main issue is:
A Plus user explicitly selecting GPT-5.6 Thinking can silently receive GPT-5.5-mini on one browser, while the same account, device, network and model setting receive GPT-5.6 correctly on another browser.
This issue has already been reported to OpenAI Support under:
Support Case #13600924
I have already provided HAR files and reproducible evidence to Support, but I have not yet received a technical explanation or permanent fix.
I can provide additional HAR files, request IDs, turn trace IDs, and further A/B test results if needed.