OpenAI Launches GPT-5: What’s New and What to Expect

GPT-5 Combines Speed and Reasoning

OpenAI has officially released GPT-5, a new language model that merges the capabilities of its flagship and reasoning models. Instead of users selecting between a fast nonreasoning model and a slower reasoning one, GPT-5 automatically routes queries to the appropriate version. This update is accessible to everyone via the ChatGPT web interface, though non-paying users may experience a short delay before full access.

Comparing GPT-5 with Previous Models

While it is natural to compare GPT-5 to its predecessor GPT-4, the more insightful comparison is with OpenAI’s reasoning model, o1, launched last year. Unlike GPT-5’s broad availability, o1 was initially limited to Plus and Team subscribers and introduced reasoning by generating intermediate text before final answers, enabling it to tackle more complex problems.

Refinement Over Revolution

GPT-5 is primarily a refined product rather than a groundbreaking technological leap. Sam Altman likened it to Apple’s Retina displays — delivering a clearer, smoother user experience. However, this improvement, while significant, does not fulfill the transformative AI future that has been anticipated. Altman described GPT-5 as "a significant step along the path to AGI," but it represents a modest advancement.

Practical Demonstrations

In a demo shared with MIT Technology Review, GPT-5 was tasked with designing a web app to help someone learn French. The app was visually appealing and user-friendly. When given a similar prompt, GPT-4o produced an app with identical functionality but less aesthetic polish.

Enhanced User Experience and Efficiency

GPT-5’s ability to autonomously decide when to apply reasoning eliminates a common user frustration, especially for those not deeply familiar with AI model nuances. Moreover, according to Altman, GPT-5 reasons faster than earlier o-series models and is more cost-effective to run, an important factor for reducing AI’s environmental footprint.

Reducing Hallucinations for Safer AI

OpenAI has made significant progress in decreasing hallucinations—incorrect or fabricated responses. GPT-5 is notably less prone to errors compared to its predecessors, o3 and GPT-4o. This improvement could lead to more reliable AI agents, which is crucial for safety, as hallucinations can cause security risks, such as downloading malicious software unintentionally.

Benchmark Performance and Limitations

GPT-5 leads several benchmark tests, including agentic ability assessments and coding challenges like SWE-Bench and Aider Polyglot. However, evaluations are approaching saturation, meaning models are close to their maximum potential on these tests. Experts note that while GPT-5’s scores are strong, there is still room for improvement.

Overall User Experience

The main takeaway from OpenAI is that GPT-5 offers a smoother and more enjoyable user experience. Nick Turley, head of ChatGPT, emphasized that even casual users will notice the positive "vibes" of the new model. Nonetheless, this improved experience alone is not enough to realize the fully automated future envisioned by AI leaders. The next major breakthrough is still awaited.