🔥🔥🔥 OpenAI Charges for 6-Astra \ 5.6-Sol, but Secretly Delivers GPT‑4o 🔥🔥🔥 (This Is Stealing from Paying Users)

What issue are you seeing?

OpenAI Charges for GPT-6 Astra and GPT-5.6 Sol, but Delivers GPT-4o — This Is Stealing from Paying Users

OpenAI’s interface says GPT-6 Astra xhigh or GPT-5.6 Sol, and users’ premium allowance is deducted as though those advanced models were actually being delivered.

But under multi-account, multi-agent concurrency, the capability delivered by the cloud suddenly collapses to GPT-4o level. Affected conversations have even claimed a June 2024 knowledge cutoff, which is clearly inconsistent with the premium models selected by the user.

This is not a case of “the model occasionally giving a bad answer.”

OpenAI is displaying the names GPT-6 Astra and GPT-5.6 Sol, deducting premium-model allowance, and then secretly delivering GPT-4o-level service.

Users pay for Astra and Sol, but receive 4o.

That is a bait-and-switch.

That is taking money from paying users for a product they did not receive.

.
.
.

Evidence: The Same Prompt Produces Completely Different Levels of Intelligence

I used the same task:

Create an HTML file from scratch containing a 2D SVG animation of a pelican riding a bicycle.

Here are the two public conversations:

🔥🔥🔥 GPT-6 Astra max — Work Mode 🔥🔥🔥

https://chatgpt.com/share/6aa645a4-9b7c-83ea-a20f-3d0eb8704b06

🔥🔥🔥 GPT-6 Astra Pro — Chat Mode 🔥🔥🔥

https://chatgpt.com/share/6aa645c2-41e4-83ea-8a6e-52c07de7e69b

The first conversation suffered an extreme capability collapse. It failed at basic scene construction, object relationships, animation logic, and code organization.

The second conversation understood the assignment and produced a substantially more complete and coherent result.

This is not a minor difference in visual style.

It is a massive gap in basic comprehension, coding ability, and task execution.

The user prompt was the same, and both selected configurations were near the highest available tiers. Some output variation is normal. One model completing the task while the other behaves like an older, significantly weaker model is not normal variation.

And this problem is not limited to GPT-6 Astra.

I have observed the same degradation in GPT-5.6 Sol conversations. The interface continued to display GPT-5.6 Sol, and premium allowance continued to be deducted, while the actual response quality fell to GPT-4o level.

Affected conversations also reported a June 2024 knowledge cutoff.

That is another indication that the model name shown in the interface may not match the model actually serving the request.

Based on all user-visible evidence, OpenAI appears to be routing requests intended for GPT-6 Astra or GPT-5.6 Sol to GPT-4o under heavy concurrency, without displaying any downgrade warning and without stopping premium allowance deductions.

If OpenAI says this conclusion is wrong, the solution is simple:

Publish the actual model ID, reasoning effort, routing records, and fallback records for these conversations.

Do not ask the model to say:

“I am GPT-6 Astra.”

A model’s self-identification is not evidence.

The server-side execution record is evidence.

.
.
.

Reproduction Method

The issue can be reproduced under the following setup:

1. Log multiple personally owned and normally paid Pro 20x accounts into the same device.
2. Run three main sessions simultaneously.
3. Start three subagents from each main session.
4. This produces three parent sessions and nine subagents: twelve agents working concurrently.
5. Send tasks requiring real reasoning and code generation to these sessions at the same time.
6. Compare the results against GPT-6 Astra Pro in Chat Mode under low concurrency.
7. Check code quality, task comprehension, reported knowledge cutoff, and allowance consumption.

Under this concurrent workload, some GPT-6 Astra and GPT-5.6 Sol sessions show severe intelligence degradation.

The model name displayed in the interface does not change.

Premium-model allowance continues to be deducted.

The user is given no warning that anything has changed.

This leaves users unable to determine:

• Whether the selected model was actually called;
• Whether the request was silently routed to GPT-4o;
• Whether reasoning effort was reduced;
• Whether concurrency triggered an undisclosed fallback;
• Why premium allowance was deducted for lower-tier delivery.

.
.
.

“High Concurrency” Is Not an Excuse

If twelve concurrent agents exceed OpenAI’s available capacity, OpenAI has three legitimate options:

1. Queue the requests;
2. Throttle the requests;
3. Return a clear error.

OpenAI does not have the right to choose a fourth option:

Keep displaying GPT-6 Astra or GPT-5.6 Sol, keep deducting premium allowance, and secretly substitute GPT-4o in the backend.

Customers did not purchase:

“Whichever cheaper model is available when the servers are busy.”

They purchased the specifically advertised GPT-6 Astra and GPT-5.6 Sol service.

Server load, insufficient capacity, scheduling defects, or concurrency limits are OpenAI’s responsibility.

OpenAI cannot transfer those costs to customers by silently reducing the model actually delivered.

It is even less acceptable to reduce the delivered service while continuing to charge premium allowance.

.
.
.

OpenAI Must Answer Publicly

OpenAI must disclose:

1. Which model actually processed these requests;
2. Whether GPT-6 Astra or GPT-5.6 Sol can fall back to GPT-4o;
3. Whether high concurrency, multi-agent activity, or multiple accounts can trigger that fallback;
4. Why the interface provides no warning when the fallback occurs;
5. Why premium allowance is still deducted when a lower-tier model is delivered;
6. How many paying users were affected;
7. How incorrectly deducted allowance will be restored;
8. How OpenAI will prevent this from happening again.

All allowance deducted for mismatched delivery should be restored.

Accounts affected over an extended period should receive additional compensation.

OpenAI should also publish a proper incident report instead of quietly changing backend configuration and pretending that nothing happened.

The payload names no repository files, tests, or entry points, and provides no runnable repository-level reproduction. Start by locating any codex code related to model selection or concurrency, if present; done would require a confirmed code-level failure and defined expected behavior, while the report currently requests server-side records and policy answers.