OpenAI published an updated system card for GPT-5.2 on 11 December 2025, documenting changes to the model's behaviour and safety profile since the original GPT-5 release [source]. The update follows user reports of inconsistent outputs and apparent capability drift in production deployments.

The system card notes that GPT-5.2 incorporates "refined post-training alignment" and adjustments to reduce what OpenAI describes as "over-refusal patterns" observed in GPT-5. Internal evaluations showed the updated model declined 12% fewer benign requests while maintaining comparable performance on adversarial prompt benchmarks.

OpenAI disclosed that GPT-5.2 exhibits measurably different behaviour on several standard tasks. Coding assistance outputs now favour brevity over verbose explanations, a shift the company attributes to user feedback. The model also shows reduced tendency to hedge statements with uncertainty markers, which OpenAI characterises as improved confidence calibration.

The update includes revised safety mitigations for jailbreak attempts identified in the wild since GPT-5's launch. OpenAI states these changes address specific prompt patterns that previously bypassed content filters, though the system card does not detail the attack vectors.

Benchmark scores remain within 2% of GPT-5 across MMLU, HumanEval, and other standard evaluations. OpenAI notes that real-world performance may vary due to the alignment changes.

The system card confirms GPT-5.2 is now the default model served via the GPT-5 API endpoint. Users accessing the API since 9 December have been interacting with the updated version. OpenAI has not announced a separate endpoint for the original GPT-5 weights.