OpenAI’s GPT-5, released in limited preview to select research partners, has demonstrated remarkable improvements in multi-step reasoning, mathematical proof generation, and scientific hypothesis formation.

Independent benchmarks conducted by Stanford’s AI Lab show the model achieving 94% accuracy on graduate-level physics problems and 89% on novel mathematical proofs — up from 67% and 52% respectively for GPT-4.

“The jump in reasoning capability is unlike anything we’ve seen between model generations,” said Dr. James Wu, who led the Stanford evaluation. “It’s not just pattern matching anymore — the model appears to construct genuine logical chains.”

However, concerns about safety and alignment have intensified. Several AI ethics organizations have called for mandatory third-party audits before wider deployment.