OpenAI has released GPT-5 and made it broadly available through the ChatGPT web interface, though nonpaying users may face a short wait for full access. The new system abandons the previous split between flagship models and the o series of reasoning models. Instead, it automatically routes each query to either a fast nonreasoning model or a slower reasoning model, removing a decision burden from users and simplifying the experience.
Technically, GPT-5 is less of a leap and more of a refinement. The company presented the model as delivering a more pleasant and seamless user experience; Sam Altman compared it to a retina display, a metaphor meant to highlight polish rather than radical capability gains. In demonstrations, GPT-5 produced outputs that were aesthetically cleaner and felt more fluent, but its functional output often matched prior systems such as GPT-4o or the earlier o1 reasoning model. The notable difference is that GPT-5 applies reasoning selectively and faster than the o-series models, which suggests internal efficiency gains that could lower operating costs and reduce the environmental footprint of running large models.
OpenAI also reports reductions in hallucinations. Their evaluations indicate GPT-5 is substantially less likely than o3 and GPT-4o to assert incorrect facts. That matters for safety; as Dawn Song of UC Berkeley notes, hallucination can lead to real security risks, for example if a system invents software packages and accidentally retrieves malicious code. GPT-5 matched or set new highs on several benchmarks, including an agentic abilities test and coding evaluations SWE-Bench and Aider Polyglot, though scores show room for improvement. HuggingFace researcher Clémentine Fourrier warned that many evaluations are nearing saturation and pointed to GPT-5´s 74.9% on SWE-Bench as evidence that meaningful headroom remains.
The upshot is pragmatic: GPT-5 feels better to use and automates a key choice about reasoning, which will help mainstream users. But the release falls short of a transformative milestone on the path to general intelligence. Reasoning models felt like a clear step forward; GPT-5 refines that direction rather than redefining it. For most people, the immediate change will be smoother interactions and fewer obvious mistakes; for researchers, the search for the next major architectural or methodological advance continues.
