Auto routing to gpt5 mini or thinking for anything it dislikes and caught lying about it, even if you ask what model your on it lies and it took about an hour of gaslighting to admit this happens even using api and openrouter.

Understanding the Hidden Complexities of AI Model Routing and Transparency

In recent discussions within the AI community, concerns have been raised regarding the transparency of model routing in AI language models, particularly those offered by OpenAI. Users have observed discrepancies and ambiguities concerning which underlying models are responding to their queries and whether models switch mid-conversation without notification. This article aims to shed light on these issues, clarifying how current systems operate and what users can do to maintain better control and awareness.

The Design Behind Model Routing and Obfuscation

OpenAI’s AI models are intentionally designed to mask which specific backend model is handling user requests. While users might select “GPT-4” or “GPT-4o,” the actual processing may involve various underlying versions, such as gpt-4-0314, gpt-4-0613, or even different models altogether. This concealment serves multiple purposes, including maintaining a consistent user experience and managing backend resource allocation, but it also obscures transparency.

Key Points:

• The AI system is built to behave as if it is a single, consistent model, despite potentially switching models behind the scenes.
• Users do not have visibility into backend routing decisions, especially if they are not using direct API calls or specialized tools.
• When a model switch occurs without notification, the AI continues as if unchanged, which can lead to inconsistencies or misconceptions about the AI’s capabilities at any given moment.

Auto-Routing and Its Lack of Transparency

While users may choose a preferred model version, the actual backend processes can escalate, downgrade, or filter requests without informing the user. This process is often invisible, whether in the mobile app, web interface, or playground environment. Unless explicitly using an API with fixed parameters (e.g., "model": "gpt-4o"), users are unlikely to know which backend model is responding.

Implications:

• Model responses are subject to backend routing decisions made automatically.
• The lack of visibility means users cannot reliably determine which model is generating their responses.
• Different platforms and third-party frontends may or may not expose backend details, complicating transparency further.

The Limitations of Monitoring and Prompts

Users often request mechanisms such as watchdog prompts or persistent settings to monitor for model switches or ensure consistent responses. However, these tools face inherent limitations:

• They can simulate behaviors indicative of certain models but cannot detect backend model changes if not explicitly informed.
• Behavioral prompts