OpenAI has released benchmark results showing that GPT-5 achieves near-human performance on complex multi-step reasoning tasks across mathematics, law, and medicine.
The model demonstrated a 94% accuracy rate on the LSAT, 91% on the USMLE Step 2, and 97% on the AMC 12 mathematics competition. These results represent significant improvements over GPT-4, which scored 88%, 83%, and 91% respectively.
Most impressively, GPT-5 showed the ability to maintain coherent reasoning chains across 15+ steps — a capability that previous models struggled with beyond 5-7 steps.
OpenAI's research paper attributes the improvement to a new training methodology called 'recursive self-refinement,' where the model iteratively improves its own reasoning during inference.
The implications for professional services are significant. Legal research, medical diagnosis, and financial analysis — all fields that require multi-step reasoning — could be substantially augmented by these capabilities.