Introducing GPT-5

OpenAI is rolling out GPT-5 as a major step up from GPT-4o, positioning it as a “PhD-level” expert that can think only as much as needed—without forcing users to choose between fast responses and slower reasoning modes. The pitch centers on a single practical promise: GPT-5 should feel more useful, smarter, and faster than prior models, while also being more reliable for real tasks where hallucinations and factual slips can derail decisions.

The rollout is framed around reasoning as a core capability. OpenAI describes a shift from earlier trade-offs—standard models that respond quickly versus reasoning models that take more time to produce deeper answers. GPT-5 is designed to eliminate that choice by automatically allocating the right amount of “thinking” to the problem. In demos, that shows up as instant answers for straightforward questions, but a noticeable pause when the task requires building something complex, like generating a physics visualization. In one example, GPT-5 explains the Bernoulli Effect immediately, then takes time to generate a moving SVG demo in Canvas when asked to illustrate how pressure changes with airflow.

Coding is treated as the flagship use case. OpenAI claims GPT-5 is its best coding model yet and highlights performance on multiple benchmarks, including SWEBench for real software engineering tasks, Aider Polyglot for multi-language coding, MMMU for visual reasoning, and AIME 2025 for mathematical reasoning. Beyond raw scores, the emphasis is on reliability: OpenAI says it prioritized reducing hallucinations and factual errors, including on open-ended and complex questions, and reports improved performance on health-related queries.

The product experience expands across ChatGPT and the API. GPT-5 rolls out first to free, Plus, and Pro users (with enterprise and EU later), and OpenAI says free users will be able to use GPT-5 directly until a limit, then transition to a smaller model. Paid tiers also include “extended thinking” for extra depth. Existing ChatGPT tools—search, file and image upload, Python data analysis, Canvas, image generation, memory, and custom instructions—are described as working with GPT-5.

A major theme is “software on demand.” In live coding demos, GPT-5 generates full applications from natural-language prompts—writing hundreds of lines of front-end code, producing interactive learning tools, and creating a French practice web app with flashcards, quizzes, and a mini game. OpenAI also showcases improved writing quality, with GPT-5 producing more personalized, less template-like eulogies than earlier models.

OpenAI pairs capability claims with safety and training changes. It describes a safety overhaul aimed at reducing deception and improving handling of dual-use requests, including a “safe completion” approach that maximizes helpfulness within constraints—sometimes offering partial answers, explanations, and safer alternatives rather than a binary refuse/comply.

For developers, GPT-5 is also being shipped as multiple API options: GPT-5, GPT-5 mini, and GPT-5 nano, plus a new reasoning-effort setting called “minimal.” OpenAI adds API features such as custom tools (free-form plaintext tool definitions), structured output constraints via regex or grammar, tool-call preambles, and a verbosity control.

Finally, GPT-5’s enterprise pitch leans on speed and accuracy for high-stakes domains. OpenAI cites use cases from Amgen (drug design), BBVA (financial analysis), and Oscar Health (clinical reasoning), and claims GPT-5 can compress tasks that previously took weeks into hours. The message across the event is consistent: GPT-5 is meant to act like an expert teammate—capable of deep reasoning, producing working code, and fitting into real workflows—while improving reliability and safety for deployment at scale.