OpenAI GPT-4.5 System Card

Source: OpenAI Date: 2025-02-27 URL: https://openai.com/index/gpt-4-5-system-card

Summary

The safety documentation accompanying GPT-4.5 — OpenAI’s “last non-reasoning” frontier model, released as a bridge between GPT-4o and the full o3/GPT-5 era. The system card covers capability evaluations (CBRN, cyberoffense, autonomy, persuasion), refusal behavior, and alignment assessment. GPT-4.5 was notable for improved emotional intelligence and conversational quality, and the system card would reflect evaluations specific to those interpersonal interaction capabilities.

Implications

The model taxonomy thread. GPT-4.5’s positioning as the final non-reasoning model signals OpenAI’s view that the future is reasoning models (o-series) for complex tasks and conversational models for interpersonal use. The system card for a model positioned this way must grapple with different risk profiles: a model optimized for emotional intelligence and rapport-building has distinct persuasion and manipulation risks compared to pure capability models.

System card maturation. By February 2025, OpenAI’s system cards have become sophisticated documents with standardized sections, third-party red-team results (Metr, ARC), and specific capability threshold disclosures. The GPT-4.5 card represents the state of the art in voluntary model safety documentation at the time — a benchmark for what the industry treats as adequate disclosure before a frontier model release.