ChatGPT Web requests GPT-5.6 Sol but server resolves GPT-5.5-mini on Plus accounts

TEB · August 22, 2026, 9:49am

Summary

ChatGPT Web intermittently routes GPT-5.6 Sol requests to GPT-5.5-mini, even though the model picker and outgoing request both specify GPT-5.6 Sol.

This is not based on asking the model which model it is. I verified the issue using ChatGPT’s network telemetry and the server-provided server_ste_metadata.model_slug field.

Environment

- Product: ChatGPT Web, not the OpenAI API

Plan: Plus
Plan metadata: plus / paid
Country: Germany
Cluster observed: germanynorth
Browsers tested: Google Chrome and Microsoft Edge
Mobile was behaviorally affected as well, but I have not captured its server metadata
First observed: August 21, 2026
The problem persisted for multiple hours

Actual result

The outgoing ChatGPT request correctly contained:

model: gpt-5-6-thinking
thinking_effort: extended

However, the server-derived telemetry for the generated response contained:

turn_analytics.server_ste_metadata.model_slug: gpt-5-5-mini
turn_analytics.server_ste_metadata.requested_model_experience: thinking
turn_analytics.server_ste_metadata.plan_type: plus
turn_analytics.server_ste_metadata.plan_type_bucket: paid

This produced the following verified mismatch:

Requested: gpt-5-6-thinking
Resolved:  gpt-5-5-mini

Expected result

When GPT-5.6 Sol is explicitly selected, the resolved model should remain GPT-5.6 Sol:

Requested: gpt-5-6-thinking
Resolved:  gpt-5-6-thinking

If a fallback is necessary, ChatGPT should clearly disclose it in the UI instead of continuing to display GPT-5.6 Sol.

Reproduction steps

Sign in to a ChatGPT Plus account.
Start a new ChatGPT Web conversation.
Explicitly select GPT-5.6 Sol with High or Extended reasoning.
Send a message.
Inspect the outgoing /backend-api/f/conversation request.
Confirm that its root model field is gpt-5-6-thinking.
Inspect the later /ces/v1/telemetry/intake request.
Compare the request model with turn_analytics.server_ste_metadata.model_slug.

During the failure, the request asks for GPT-5.6 Thinking but the server metadata reports GPT-5.5-mini.

Additional metadata

Conversation metadata reported:

default_model_slug: gpt-5-6-thinking
blocked_features: []
model_limits: []
workspace_model_policy: null

This does not look like an advertised model limit, blocked feature, workspace policy, or incorrect picker selection.

Browser/session-dependent behavior

Clearing all ChatGPT site data in Microsoft Edge and signing in again created a working session.

The same account then produced:

Requested: gpt-5-6-thinking
Resolved:  gpt-5-6-thinking
Result:    Match

This suggests the failure may depend on browser identity, session state, device assignment, OBI, feature flags, or a server-side routing cohort rather than an account-wide entitlement.

The affected session also showed possibly related identity/bootstrap errors:

OBI synchronization is not available
bazaar_personalization_unavailable: true
persona: chatgpt-noauth

The interface occasionally appeared logged out and returned to the authenticated state after a page refresh.

Impact

The UI continues to show GPT-5.6 Sol while the server metadata identifies GPT-5.5-mini. Response latency, reasoning behavior, instruction following, and output quality differ substantially.

This makes it impossible for users to know whether their selected paid model is actually serving the response without inspecting internal telemetry.

Request

Could OpenAI investigate:

Why gpt-5-6-thinking requests are resolved as gpt-5-5-mini.
Whether routing is incorrectly keyed to browser, session, device ID, OBI, region, or an experiment cohort.
Whether this is related to the simultaneous authentication/OBI synchronization problems.
Why the UI does not disclose the resolved fallback model.
Whether affected sessions can be reset server-side.

I can provide UTC timestamps and request IDs privately to OpenAI staff. I will not post authentication headers, cookies, session tokens, device IDs, OBI values, conversation IDs, or a full HAR publicly.

a19749893 · August 22, 2026, 4:05pm

I’m also experiencing this issue on my side. It seems that this behavior is affecting more than one user. Sharing additional examples in the main thread may help provide a clearer picture of what is happening.

krasnik · August 23, 2026, 11:31am

I have an independent Plus-account result that may help narrow this down.

I was also seeing a GPT-5.6 Thinking request resolve incorrectly in one browser environment. Instead of deleting chats or changing account settings, I tested the same account in a completely separate fresh Chromium browser environment.

In my case, the clean environment restored genuine GPT-5.6 Thinking routing.

The important detail is that this was stronger isolation than simply opening Incognito or adding another normal browser profile. The working environment used a completely separate Chromium user-data directory, with no imported browser data, no Sync, and no pre-existing extensions or ChatGPT state.

So the comparison was effectively:

Affected existing browser/session → incorrect routing
Same Plus account, fresh isolated browser environment → GPT-5.6 Thinking restored

This seems consistent with your observation that clearing Edge state and signing in again produced a working session.

I would be careful about concluding that “old cookies” themselves are the root cause, though. The evidence only shows that changing the browser/session environment can change the routing outcome. A fresh browser identity/session may also cause different server-side treatment, cohort assignment, or routing state.

One other useful control from my case: after moving to the clean environment, I verified that the model-routing problem was gone, but a separate long-running execution/runtime regression remained. So I would currently treat the 5.5-mini routing problem and broader Thinking/runtime issues as potentially distinct rather than assuming they share one cause.

If other affected users can reproduce an old environment broken / fresh isolated environment working split on the same account, that would be a very useful pattern to collect.

DLPB · August 23, 2026, 12:48pm

See my latest post here Connection issues on top of 5.6 thinking on high and extra high regression - #7 by alithebigg1

TEB · August 23, 2026, 1:19pm

Thanks, i tried but it didnt fix it unfortunately, i also bought a fresh account but luckily that aswell did not work. Seems like i just gotta hope OpenAI support becomes useful

momoto · August 23, 2026, 4:56pm

Confirming the same issue on a different cluster.

Plan: Plus
cluster_region: westus3
Requested: gpt-5-6-thinking
default_model_slug: gpt-5-6-thinking
resolved_model_slug: gpt-5-5-mini

Onset: Aug 20. Unlike the intermittent behavior described in the OP, this is 100% consistent on my account — every single message resolves to mini, including the first message of a brand new conversation.

Verified via network telemetry (streamed metadata), not by asking the model to self-identify.
Tried: new conversations, cleared cookies, re-login. No effect.

Two data points worth noting:

OP is on germanynorth, I’m on westus3 — this does not appear to be a single-cluster failure.
Reasoning effort selection has no observable effect; all levels resolve identically.

Happy to provide additional captures if useful.

finley.lee · August 25, 2026, 7:16am

Hello OpenAI team and community,

I would like to report a possible GPT-5.6 Sol routing issue affecting my ChatGPT Plus account.

This does not appear to be a simple quality difference. It looks like a model routing / fallback transparency issue.

Environment:

Subscription: ChatGPT Plus
Platforms tested:
- Web
- iOS app
- Android app
Chat mode: Regular Chat (not Work mode / workspace)

Observed behavior:

The same account behaves differently depending on the reasoning level:

GPT-5.6 Sol with normal / no deep reasoning:

Works normally.

GPT-5.6 Sol Fast reasoning:

Works normally.

GPT-5.6 Sol Medium reasoning:

Responses appear to be handled by GPT-5.5-mini.

GPT-5.6 Sol High reasoning:

Responses appear to be handled by GPT-5.5-mini.

This happens on the same account and during the same period of time.

Important details:

GPT-5.6 Sol remains selectable in the model picker.
Medium and High reasoning options remain available.
I do not receive any clear warning that a reasoning limit has been reached.
I do not receive any reset-time information.
There is no clear notification that the conversation has been moved to another model.

I understand that fallback mechanisms may exist for capacity management.

However, the concern is transparency:

If a user selects GPT-5.6 Sol Medium/High, but the request is actually served by GPT-5.5-mini, users should receive a clear notification.

For paid subscribers, the selected model and actual serving model should not silently differ.

I have already contacted OpenAI Support, but the AI-assisted support response cannot verify backend routing. I am hoping someone from the community or OpenAI team can help clarify whether this is:

An expected GPT-5.6 reasoning limit fallback;
An account/session routing bug;
A UI/model labeling issue;
Another backend routing problem.

I can provide:

screenshots;
affected conversation links;
timestamps;
device/app information.

Has anyone else experienced this pattern?

Especially:

GPT-5.6 Fast works;
GPT-5.6 Medium/High falls back to GPT-5.5-mini.

Thank you.

finley.lee · August 25, 2026, 11:46am

Subject: FINAL ESCALATION — Serious Trust and Transparency Failure Regarding GPT-5.6 Routing Issue

Dear OpenAI Support Team,

I am writing this as a final escalation regarding my GPT-5.6 routing issue.

I want to express my deep disappointment, frustration, and concern regarding both the issue itself and the way this matter has been handled.

I am not writing as a casual user.

I am a long-term ChatGPT Plus subscriber and a highly engaged user who has genuinely supported OpenAI. I have integrated ChatGPT deeply into my professional workflow, including research, complex analysis, writing, document processing, spreadsheet work, and long-context problem solving.

I have followed OpenAI’s development closely and have always considered ChatGPT one of the most valuable AI products available.

As a user from China, maintaining stable access to international AI services is not always easy. The additional technical difficulties involved in accessing and using these services require extra effort and commitment. Despite those difficulties, I continued using and supporting ChatGPT because I truly believed in this product.

That is why this experience has been extremely disappointing.

The issue I reported

This is not simply a complaint about model quality.

This is a serious concern regarding model routing transparency.

I selected GPT-5.6 Sol and tested the behavior across multiple platforms:

ChatGPT Web;
iOS app;
Android app.

My testing showed a very specific pattern:

GPT-5.6 Sol Fast / normal reasoning appears to work normally;
GPT-5.6 Sol Medium appears to behave differently;
GPT-5.6 Sol High appears to behave differently.

In particular, Medium and High reasoning responses appear to fall back to GPT-5.5-mini, while Fast/non-deep reasoning continues to behave as GPT-5.6.

This occurred on the same account and during the same period.

I have already provided:

affected conversation links;
screenshots;
screen recordings;
timestamps;
timezone information;
device and platform information.

Affected sessions:

GPT-5.6 Sol High:

GPT-5.6 Sol Medium:

GPT-5.6 Sol Low:

I have also provided iOS screen recordings as requested.

I want to be clear:

I am not a backend engineer.

I do not have access to OpenAI’s internal routing logs, server metadata, or account-level system information.

However, I have provided everything that a normal paying user can reasonably provide.

Only OpenAI has the ability to verify what actually happened in the backend.

The core problem is transparency

I understand that AI systems require capacity management.

I understand that rate limits and fallback mechanisms may exist.

That is not the issue.

The serious issue is this:

A paying user selects a premium model and reasoning level, but there is no clear explanation when the actual experience appears different.

Users need to know:

Is GPT-5.6 actually being used?
Has a reasoning limit been reached?
Has the request been routed to another model?
Is this expected behavior or a technical issue?

A premium AI service depends on trust.

If users cannot confidently understand what model is handling their work, that trust is damaged.

This does not appear to be an isolated incident

I would also like to point out that similar concerns have been raised by other users.

Examples include discussions in public communities:

OpenAI Community:

“Possible GPT-5.6 Sol routing mismatch: selected GPT-5.6 but server metadata shows GPT-5.5-mini”
“GPT 5.6 rerouting to 5.5 mini”

Reddit discussions:

“GPT-5.6 Sol users being routed to GPT-5.5-mini instead?”
“GPT-5.6 Sol High is still silently falling back to 5.5-mini”

I am not claiming these reports alone prove the technical cause.

However, when multiple users report similar behavior, this should be treated as a potential product-level issue rather than only an individual account misunderstanding.

I strongly encourage OpenAI to investigate whether there is a broader issue involving:

routing logic;
reasoning-level limits;
fallback transparency;
model selection accuracy.

My disappointment with the support experience

What has disappointed me most is not only the technical issue.

It is the difficulty of obtaining a clear answer.

I understand that AI-assisted support can help handle common questions.

However, when a user reports a possible model-routing issue that requires backend verification, repeatedly receiving general explanations about fallback policies does not resolve the actual concern.

The question has always been simple:

What happened to my account and my sessions?

I am not asking for unlimited usage.

I am asking for transparency and a proper investigation.

Impact on my subscription and trust

I need to clearly state the impact this incident has had.

I have already cancelled my ChatGPT Plus subscription.

My subscription is scheduled to end on:

August 31, 2026.

This was not an easy decision.

I did not cancel because I dislike ChatGPT.

In fact, the opposite is true.

I cancelled because I valued ChatGPT enough to expect a higher standard.

Before this incident, I was seriously considering upgrading to the $100/month Pro plan.

Instead, this experience has created a serious crisis of trust.

Loss of confidence and consideration of alternatives

I want OpenAI to understand the long-term impact of this incident.

For a professional AI service, technical problems can happen.

No complex system is perfect.

However, trust is built through:

transparency;
accountability;
reliable communication.

This incident has made me question whether I can continue relying on ChatGPT as a critical professional tool.

Because of this loss of confidence, I am now actively evaluating alternative AI products, including:

Claude;
Grok;
other competing AI services.

I may move my professional AI workflows away from ChatGPT.

I want to emphasize:

This is not because I wanted to leave OpenAI.

It is because this experience has changed my confidence in the reliability and transparency of the service.

Before this happened, I was not looking for alternatives.

I was planning to invest more in OpenAI.

Now, I am considering leaving.

Perhaps my $20/month subscription is insignificant from OpenAI’s perspective.

However, every long-term user represents trust, loyalty, and future growth.

Losing a user is not only about losing a monthly payment.

It is about losing confidence from someone who believed in the product and was willing to invest more.

Final request

I respectfully request that this case receive a final human technical review.

Please investigate:

Whether my GPT-5.6 Medium/High requests were routed to another model;
Whether my account reached a GPT-5.6 reasoning allowance;
Whether this behavior is expected, temporary, or caused by a technical issue;
Whether other users are experiencing the same behavior;
Why users are not clearly informed when fallback occurs.

Please do not treat this as another ordinary support request.

This is a formal complaint from a paying customer whose trust in the product has been seriously affected.

I still appreciate ChatGPT.

I still hope OpenAI can address this properly.

If OpenAI provides a transparent explanation and resolves this issue, I am willing to reconsider my decision.

But without a clear explanation and meaningful action, my subscription will end on August 31, 2026, and I will move forward with alternative AI solutions.

I hope OpenAI understands that this is not merely about one user or one subscription.

It is about whether users can continue trusting the relationship between themselves and the AI service they pay for.

I look forward to a substantive response from a human specialist or technical team.

Sincerely,

Finley

ChatGPT Plus Subscriber

tanguo081023 · August 25, 2026, 7:12pm

Has your issue been resolved?
I ran into the same issue. It started with Codex — I thought it was a network problem on my end, so I tried connecting from a different location, but then the same thing happened on both the web and mobile app, where requests were being routed to gpt-5.5-mini.

finley.lee · August 25, 2026, 7:30pm

Not resolved yet. I’ve been experiencing this issue since August 24. I select GPT‑5.6 Sol with Medium/High reasoning, but the responses are clearly at the level of GPT‑5.5‑mini—fast but with severely degraded quality. I’ve sent multiple emails to OpenAI Support, attaching screenshots, screen recordings, conversation links, timestamps, and test results across platforms (Web, iOS, Android). However, the support staff have not conducted any actual backend investigation on my account; they just keep asking for more screen recordings, repeatedly stalling. I’ve spent a lot of time cooperating, but I’ve never received any substantive explanation or fix. Completely disappointed by the lack of transparency and the way this has been handled, I’ve cancelled my ChatGPT Plus subscription and initiated a refund request through Google Play, and I’ve formally notified the support team. The issue still persists, and I’m still waiting for a real answer.

finley.lee · August 25, 2026, 8:06pm

Dear OpenAI Leadership and Support Team,

I am writing this message because my disappointment with OpenAI has reached a level that I never expected as a long-term paying customer.

I want to be completely clear: this is no longer only about a technical issue. This has become a serious question about whether OpenAI still respects the users who pay for and support its products.

I am a ChatGPT Plus subscriber. Like many other paid users, I chose to support OpenAI because I believed this company represented the future of artificial intelligence and that it valued transparency, reliability, and its relationship with customers.

However, my recent experience has seriously challenged that belief.

I selected a high-capability model as a paying user. Yet the actual system behavior strongly indicated that I may have received a different, lower-capability model without clear disclosure. After reporting this issue, I provided everything requested by your team, including recordings, conversation links, and detailed evidence.

What has been most disappointing is not only the possible technical problem itself.

The most disappointing part is OpenAI’s response.

A company that charges users for premium access should not leave paying customers to repeatedly prove that a problem exists while receiving no clear explanation. Users should not have to investigate the system themselves, analyze responses, record evidence, and fight for basic transparency.

This creates a deeply troubling impression:

Does OpenAI still value Plus and Pro users?

Or has OpenAI reached a point where it believes paying customers are replaceable and no longer deserve meaningful communication?

If OpenAI’s strategy is to build the future of AI, then customer trust cannot be treated as an insignificant support issue. The users who pay, test, provide feedback, and rely on your products in their professional work are not just sources of revenue. They are the foundation of your product ecosystem.

A premium subscription creates an expectation:

The service should be transparent.
The capabilities provided should match what is represented.
Problems affecting user trust should receive serious investigation.
Paying customers should receive clear and respectful communication.

At this moment, I do not feel that these expectations have been met.

This incident has already affected my work, my productivity, my sleep, and my confidence in relying on ChatGPT as a professional tool. The amount of time and energy I have spent trying to obtain a reasonable explanation has been unacceptable.

I want to emphasize that I appreciate individual support representatives who attempted to assist me. However, the overall response from OpenAI has given me the impression that the company does not fully recognize the seriousness of this situation.

A company can survive technical failures.

A company cannot easily recover from losing user trust.

Therefore, I am requesting a formal response from a responsible OpenAI representative, not a generic customer support message.

I expect OpenAI to address the following:

Whether paid users can trust that the selected model is actually the model being provided.
Whether undocumented routing, fallback, or capability changes are occurring.
Why affected users were not proactively informed.
What actions OpenAI will take to restore confidence among paying customers.

I sincerely hope OpenAI understands the significance of this situation.

The future of AI will not be determined only by model capability. It will also be determined by whether users believe the companies building these systems are honest, transparent, and worthy of trust.

At present, my trust in OpenAI has been seriously damaged.

I expect a serious response.