Router Misrouting: The system often uses a faster, less intelligent sub-model for your query.
Chat vs. API Discrepancy: The web chat gives different results than the developer API because it uses an automatic router.
Model Drift: Your old, reliable prompts may no longer work as well because the underlying model has changed.
Long Context Illusion: The model can forget information from the middle of a very long document, despite its large context window.
Inconsistent JSON Output: The model sometimes fails to generate valid, properly formatted JSON code.
Hallucinated Tool Actions: The model may falsely claim it used a tool (like running code) when it actually didn't.
Empathy and Speed Issues: The default model can feel robotic and lack empathy because it's optimized for speed, while the "Thinking" mode can be too slow.
Overly Cautious Guardrails: The model's safety filters can be overly sensitive and refuse valid requests, especially in scientific fields.
Basic Errors: The model sometimes makes simple, basic errors on tasks it should handle easily.
The Silent Fallback: After many messages, the model may silently switch to a lower-quality version to manage usage limits without telling you,