OpenAI has finally released GPT-5. The new system does away with the distinction between OpenAI flagship models and its O series of reasoning models. Instead, it automatically routes user queries to either a faster nonreasoning version or a slower reasoning model. The new system is now available through the ChatGPT interface. However, nonpaying users will need to wait for a few days before they can access the full capabilities.
GPT-5 is tempting to compare with its explicit predecessor GPT-4. However, a more insightful comparison would be with o1, OpenAI’s first reasoning model released last year. Unlike GPT-5, which was released to the public, o1 initially only available to Plus and Team members. These users had access to a new language model that “reasoned” through its responses by generating more text before providing a response. This allowed it to solve far more difficult problems than its nonreasoning equivalents.
GPT-5 is a refined version of o1, which was a major technological advance. Sam Altman, during a press conference, compared GPT-5 with Apple’s Retina display. It’s a good analogy, but perhaps not the one he intended. GPT-5, like a screen with unprecedented clarity, will provide a smoother and more enjoyable user experience. Altman has spent the last year hyping the AI future he envisions. Altman said that GPT-5 was “a significant step on the path to AGI,” which is artificial general intelligence. He may be right, but it’s still a very small one.
OpenAI has shown a demo of the model’s capabilities. MIT Technology Review In advance of its release. Yann Dubois asked GPT-5, a post training lead at OpenAI to design a website application that would help her learn French, so that she could better communicate with his family. The model followed his instructions admirably and created a user-friendly, appealing app. When I gave GPT-4o a prompt that was almost identical, it produced a similar app. The only difference was that it wasn’t quite as attractive.
Some of the other improvements to user experience are more substantial. The model, rather than the user, choosing whether to apply reasoning to every query removes an important pain point. This is especially true for users who do not follow LLM developments closely.
Altman says that GPT-5 is much faster at reasoning than the o series models. OpenAI’s decision to release it to non-paying users indicates that it is also less expensive to run. This is a big deal. Running powerful models quickly and cheaply is a difficult problem. Solving it is the key to reducing AI’s environmental impact.
OpenAI has taken steps to reduce hallucinations. This has been a persistent problem. OpenAI’s evaluations indicate that GPT-5 models make significantly fewer incorrect claims than the predecessor models, GPT-4o and o3. If this advancement is able to withstand scrutiny, it may pave the road for more reliable and trust-worthy agents. Dawn Song, a professor at UC Berkeley who teaches computer science, says that hallucination can lead to real safety and security problems. A hallucinating agent could, for example, download malicious code onto a user’s computer.
GPT-5 is the best in the world on several benchmarks. These include a test for agentic abilities, the SWE-Bench coding evaluation and Aider Polyglot. Clementine Fourrier is an AI researcher with the company HuggingFace. She says that these evaluations are close to saturation, meaning that current models have reached near-maximum performance.
She says that it’s like comparing the performance of high school students on middle-grade problems. “If the high-schooler fails, you learn something, but if they succeed, you don’t learn much.” Fourrier said that she would have been impressed if SWE-Bench had scored 80% or 85%, but it only managed to score 74.9%.
OpenAI’s main message is that GPT-5 is easier to use. “The vibes are really good and I think people will really feel that, especially the average people who haven’t spent their time thinking about model,” said Nick Turley.
Altman’s automated future will not be achieved by just using vibrations. The ability to reason felt like a big step forward in the journey towards AGI. We’re still awaiting the next one.
