OpenAI GPT-4.5 System Card

Summary

System card for GPT-4.5, OpenAI's Feb 2025 research-preview chat model: scales unsupervised pretraining on top of GPT-4o (SFT + RLHF plus new supervision) as a general-purpose alternative to STEM reasoning models. Claims a more natural interaction, broader knowledge, stronger intent following, higher EQ, and fewer hallucinations; positioned for writing, programming, and practical problem-solving rather than new frontier capabilities. Preparedness scorecard post-mitigation: CBRN medium, persuasion medium, cybersecurity low, model autonomy low; overall medium (deployable). Safety evals cover disallowed content, jailbreaks, mistakes, and the four Preparedness categories. Cite as OpenAI (2025). Later retired from ChatGPT (sunset announced for 2026-06-27).

Keywords

blog · gpt-4.5 · system-card · openai · safety · preparedness · cbrn · persuasion

Topics

foundation models, safety, system cards

Research notes

• Primary: posted PDF https://cdn.openai.com/gpt-4-5-system-card.pdf. HTML index https://openai.com/index/gpt-4-5-system-card/. Alternate CDN filename gpt-4-5-system-card-2272025.pdf also circulates. Closed model; no weights. Not a hosted corpus, so no datasets_local row.