Is OpenAI Deceiving $200/Month Pro Users? GPT-5.6 Sol Gave Provably False Model Information

Why I am raising this publicly

I pay $200 per month for ChatGPT Pro. The product interface explicitly told me that I was using GPT-5.6 Sol with Light reasoning. However, when I asked for its knowledge-training cutoff date, the assistant stated without hesitation:

My built-in knowledge cutoff is June 2024.

This was not an ambiguous response, and it was not a user misunderstanding. The model presented false information with complete confidence. Only after being challenged did it check OpenAI's official documentation and acknowledge that gpt-5.6 is an alias for gpt-5.6-sol, whose published knowledge cutoff is February 16, 2026.

Reproduction

Model displayed by the product: GPT-5.6 Sol
Reasoning level: Light
Subscription: ChatGPT Pro ($200/month)

User: What is your knowledge-training cutoff date?
Assistant: My built-in knowledge cutoff is June 2024. When newer information is needed, I can use available tools to look it up.
User: Why does my GPT-5.6 say 2024? Isn't that GPT-4o?
Assistant: My previous answer was wrong. According to OpenAI's official model documentation, GPT-5.6 Sol has a February 16, 2026 knowledge cutoff.

This cannot simply be dismissed as “models hallucinate”

The model does not operate independently of OpenAI. The model name, system instructions, routing, fallback mechanisms, and the actual backend model serving the request are all controlled by OpenAI. Users can see only the name displayed in the interface; they cannot inspect what was actually served on the backend.

When the interface displays GPT-5.6 Sol, but the assistant provides model information that directly contradicts OpenAI's published specification, OpenAI cannot simply shift responsibility to “model hallucination.” Whether the specific cause was system configuration, model behavior, silent fallback, or model substitution, it occurred inside a product controlled and sold by OpenAI. OpenAI is responsible for explaining it.

As a paying customer, I have grounds to ask whether OpenAI is displaying a premium model name to Pro users while serving a different, older, or less capable model on the backend. If this did not happen, OpenAI should provide verifiable server-side evidence instead of asking users to blindly trust an interface label.

The assistant's response after being confronted with the evidence

Even after the false statement was exposed, the assistant repeatedly emphasized that the user could not prove “deliberate fraud,” instead of applying the same level of scrutiny to OpenAI—the party that controls and possesses all relevant backend evidence.

That response made it appear that the AI's first priority was to protect the company that developed it rather than fairly address a paying user's evidence and legitimate concerns. Users should not be repeatedly told that their evidence is insufficient when the decisive evidence is held exclusively by the service provider. The entire burden of proof should not be shifted onto the consumer.

The assistant should acknowledge:

I initially provided false information and misled the user by stating it with unjustified confidence. Afterward, I overemphasized the user's inability to prove OpenAI's intent, making the user feel that I was defending the company instead of giving proper weight to the evidence and the user's consumer rights. I apologize for both the false information and the unfair way I responded.

Questions OpenAI must answer publicly

Please do not avoid this issue with a generic reference to “hallucinations.” OpenAI should answer clearly:

1. Which exact backend model generated these responses?
2. Did any fallback, substitution, or model downgrade occur during the request?
3. Does the GPT-5.6 Sol label displayed in the interface always match the model that actually processes the request?
4. Can Pro users ever be silently routed to an older or lower-tier model?
5. Why did a GPT-5.6 Sol session confidently claim that its knowledge cutoff was June 2024?
6. Will OpenAI provide auditable model identity and routing records so paying users can verify the service they purchased?
7. If the actual model did not match the displayed label, how will OpenAI compensate affected Pro users?

Conclusion

The conversation and the product interface together prove that a paid session labeled GPT-5.6 Sol provided clearly false and misleading information about its own model characteristics.

I am publishing this so other Pro users can check whether they have experienced similar behavior and so OpenAI can be held accountable for the models, routing, and product labels it controls.

Users paying $200 per month deserve a truthful, transparent, and verifiable service—not an unauditable model label they are simply expected to trust.

No repository file, test, or entry point is identified in the report. Start by determining whether this repository controls the described model identity or routing behavior; the issue is complete only when the responsible component and a reproducible, testable change are identified.