GPT-5 is here, and OpenAI is being unusually direct about what it can do — including reasoning that finally clears the bar set by its loudest critics.
OpenAI released GPT-5 to developers and consumers on Wednesday, kicking off the next round of "is this AGI yet" debates with a model that, by the company's own benchmarks, lands a step-change ahead of GPT-4 across most reasoning tasks.
The headline upgrades: long-horizon planning, better tool use out of the box, and a multimodal stack that handles documents, images, audio, and video without the usual seam lines. Early independent evaluations match the company's framing on the easy benchmarks; the harder ones will take weeks to settle.
For agents, GPT-5 is a noticeable level-up. Several agent platforms — including the wave that recently piped models into trading and research workflows — pushed updates within hours. Latency and cost both came down, which matters more than most users realize.
The harder questions, of course, remain. The model is still confidently wrong on a non-trivial percentage of niche queries, and "agentic" still has a long way to go before it is genuinely safe to leave running unsupervised. But the curve keeps bending, and competitors are visibly playing catch-up.