A one-question test for whether your model has been quietly downgraded
Set reasoning effort to Low and ask for the knowledge cutoff date; a mismatched answer suggests the model has been swapped out.
To verify whether your gpt-6-astra instance in Codex has suffered silent model degradation, configure the reasoning effort to Low (light) and query its knowledge cutoff date. A response indicating June 2024 strongly suggests the model has been degraded or routed to a fallback version, though it is not definitive. You can proceed with a second benchmark prompt: "Create an HTML file with an animated 2D SVG of a pelican riding a bicycle." (retaining the astra light configuration). If the generated output matches the style in Figure 1, degradation is confirmed; an output resembling Figure 2 indicates normal performance. However, even if normal initially, subsequent queries may still trigger silent throttling or degradation on the account. This inconsistency poses significant risk, as previous regressions have already corrupted my project codebase. Consequently, I must periodically execute this "pelican on a bicycle" benchmark (consuming approximately 10% of the 5-hour Plus quota). In its current state, Codex is practically non-viable for production workflows. Although @thsottiaux reported deploying a fix today, recent test runs confirm the issue persists entirely. @sama @ChatGPT @OpenAI @Trinkle23897 @TairanHe99 #chatgpt #openai #Astra #gpt6