OpenAI introduced GPT-5 on Thursday, a model that the vendor described as being its most sophisticated AI system yet and useful for coding and writing tasks.
OpenAI said GPT-5 is not just a model, it's also a system with a deeper reasoning model than its previous version. The system has a real-time router that can select what type of model to use based on the conversation. Along with fewer hallucinations and the ability to better follow instructions, GPT-5 can create websites, apps and games with one natural language prompt, according to OpenAI. It can also better handle language with structural ambiguity, such as free verse. These improvements mean ChatGPT can help with everyday tasks such as drafting and editing reports, emails and memos.
GPT-5 is also good at visual, video-based and scientific reasoning, OpenAI said. The vendor said the model performs better at thinking than OpenAI o3.
GPT-5 was unveiled two days after OpenAI released two new open-weight models. It's available to all users, but on a limited basis to those who don't pay for the service.
There's a lot of anticipation for this.
Arun ChandrasekaranAnalyst, Gartner
Despite these improvements, industry experts called GPT-5 an incremental release.
"There's a lot of anticipation for this," said Arun Chandrasekaran, an analyst at Gartner. "But [It's] nowhere close to AGI [artificial general intelligence]."
A user's take
The distributed file share service Box, an early user of GPT-5, performed an evaluation of GPT-5 compared with GPT-4.1 and found that the model is better at dissecting complicated documents in sectors such as finance. GPT-5 is now available in Box AI Studio for Enterprise Advanced customers and with Box AI APIs.
"For anything that has a lot of math and complicated information, like contract financial documents, this is where we see a healthy improvement," Aaron Levie, CEO of Box, said in an interview with Informa TechTarget.
He added that when it comes to hallucinations, Box saw a better improvement with GPT-5 than with GPT-4.1 and other models from other providers.
"Where GPT-4.1 and other models would sort of answer a question, trying to deduce information that might be in a document -- which is a little bit of a hallucination -- GPT-5 would come back and say, 'This information is not actually in the document, so I can't answer the question properly,'" Levie said. "That's a good way to identify what kind of hallucination tends to be happening."
He said the model showed a 90% accuracy rate on enterprise data documents, but there's room for improvement.
"There's still a gap where more improvements in the model will have improved accuracy over time," Levie continued. "We just want that number to go up more and more."
A focus on coding
While GPT-5 shows the improvement in accuracy, performance and safety, it also indicates that OpenAI sees an opportunity in software engineering due to how it improved the model's coding performance, according to Chandrasekaran.
"It looks like a lot of the pre-training and post-training went to the coding," Chandrasekaran said. "Coding-related tasks, particularly from a B2B perspective, are one of the fastest-growing use cases for Gen AI today."
He added that coding has been a domain in which Anthropic, an OpenAI competitor, has found success. Anthropic released its Claude Opus 4.1 model Tuesday, which improves its research and data analysis skills, according to the company.
"For OpenAI, it needs to establish itself as a leader in that space because it works with a lot of ISPs[internet service providers] that use OpenAI models underneath the covers, but I think some of them have also possibly Anthropic's models," Chandrasekaran continued.
It's unclear, though, how GPT-5 compares with other GPT models in different domains.
"Are we assuming that the model just by and far is going to be so much better at everything?" Chandrasekaran said. "Because you're making so much progress in one specific domain, are you going to take a little of a backseat in maybe another domain? We don't know that yet."
Bradley Shimmin, an analyst at The Futurum Group, said being more domain-specific is especially important for enterprise customers.
"We're starting to get to a place in the AI market where customers should expect to invest in models built to support whatever it is they're building, not just [be] general purpose," Shimmin said. "You want the option as an enterprise developer to build a solution tailored to whatever domain you're working in."
Optimization and enterprise customers
Shimmin added that OpenAI is focused on optimization by allowing GPT-5 to choose the right model for the right question.
For enterprise customers, interest in the new model will come down to the use case, according to Shimmin. He said that an enterprise developer who might need to put an entire codebase into a model will likely choose Gemini 2.5 due to its bigger context window. In contrast, a developer already in the OpenAI or Microsoft ecosystem might select GPT-5.
"It all comes down to inertia and technical debt," Shimmin said. "If I'm building software, what I've invested in yesterday is going to drive my decisions for tomorrow."
Chandrasekaran mentioned that it seems unclear if OpenAI made any improvements in any language other than English. Moreover, the vendor missed an opportunity to show how it will differentiate its proprietary models from the open weight models it just released.
"When will one be better than the other?" Chandrasekaran added. "How will they handle capacity and capacity constraints that they face? These are some things we need to watch out for in the next few weeks."
Editor's note: Informa TechTarget is a Box customer.
Esther Shittu is an Informa TechTarget news writer and podcast host covering artificial intelligence software and systems.