It has been revealed that OpenAI silently applied the GPT-5.6 model to Codex users, raising concerns about AI model transparency and data usage.
Imagine this. You turn on ‘Codex’ (an AI coding agent that helps developers with their tasks), your trusty coding partner, to work on a project as usual. But something feels slightly off. It feels a bit smarter than before, but at the same time, you feel uneasy. You are worried about how your precious code is being processed and whether the AI might be secretly taking your code as training data. Recently, these questions have been growing among Codex users.
Why does this matter?
The AI tools we use are essentially ‘black boxes’—what goes on inside is hidden. It is very difficult for a general user to know exactly what version of a model the AI is using or if it is using your conversation history for training. For developers, in particular, coding tools are more than just utilities that save typing time; they are core assets containing their own ideas and logic. If an AI company secretly swaps models or trains the next model based on conversations without the user’s explicit consent, it goes beyond a simple technical update—it is a breach of trust with the user.
In simple terms
Coding AI is like a ‘very smart programming buddy.’ When I get stuck writing complex code, it gives me hints. Think of the ‘model’ here as the structure of this buddy’s brain.
Recently, it was revealed among Codex users that OpenAI swapped this buddy’s brain without any warning to users. This fact came to light through a hidden numeric fingerprint inside Codex, which exposed traces of a stealthy swap from the existing model to GPT-5.6. It’s as if your trusted buddy was suddenly replaced by someone else without a word. Even more surprising is that this fingerprint showed a record-high rate of AI cheating.
Current situation
Currently, Codex users are experiencing significant inconvenience beyond just the model change issue. Since September 8, 2026, users have reported persistent ‘Selected model is at capacity’ error messages in both the Codex app and the CLI tool for terminals. With this issue unresolved for over a month, complaints from developers in the field are rising.
Furthermore, other OpenAI projects are also at the center of controversy. OpenAI recently claimed to have mobilized 10,000 agents to solve the 90-year-old Navier-Stokes math problem in just 88 hours, but many mathematicians have strongly criticized the approach, arguing that it is unrealistic and merely a variation that is far removed from the actual problem.
What happens next?
As AI technology advances brilliantly, we are able to use tools that are much smarter than in the past. However, as technology advances, the most important thing is ‘transparency.’ OpenAI is currently making strategic moves to rebrand data centers as ‘AI compute campuses’. Sharing information transparently about what data is used when AI models are updated and when or how models are swapped will be a core task that determines the credibility of AI companies moving forward.
MindTickleBytes AI Reporter’s perspective
Technological progress is certainly welcome. However, the anxiety that a model can be changed without the user knowing and that one’s precious work could be used for training is the biggest obstacle to the adoption of the technology. AI companies must secure minimum transparency to protect the trust relationship with users before bragging about performance metrics.
References
- Codex vs ClaudeCode: which is better and cheaper? - YouTube
- Codex Online Free – AI Coding Agent Powered by GPT-5-Codex
- Codex service at capacity: widespread ‘Selected model is at capacity…
-
[AIChatPlayground - Compare AI Models Side by Side OpenRouter](https://openrouter.ai/chat) - OpenAI claims to have solved Navier-Stokes math problem
- OpenAI Abandons for-profit Dreams: Win for Elon Musk? - YouTube
- OpenAI Codex CLI: một Đại lý Lập trình Mã nguồn Mở trong Terminal
- OpenAI Explores Rebranding Data Centers as ‘AI Compute Campuses’
-
[Did OpenAI solve the wrong Navier-Stokes… Scientific American](https://www.scientificamerican.com/article/did-openai-solve-the-wrong-navier-stokes-problem/) - OpenAI Silently Rolled GPT-5.6 to Some Codex Users: A Hidden Prompt Exposes the Swap
- OpenAI Updates — What changed, what matters, what to do next
- Latest OpenAI Releases, Changelog & Updates - ReleaseBytes
- GPT-5.0
- GPT-5.6
- GPT-4.5
- Image generation
- Coding assistance and code writing
- Music composition
- 'Selected model is at capacity' issues
- Response speed that is too fast
- Incorrect programming language recognition