Introducing GPT-5

OpenAI’s launch of GPT-5 marks a significant evolution from previous models, positioning it as a “legitimate PhD-level expert” available on demand. The central theme is a leap in reasoning capabilities, enabling the model to function not just as a knowledge retriever but as a powerful agent that can perform complex, multi-step tasks, most notably in software development.

A New Reasoning Paradigm

A core innovation in GPT-5 is its enhanced reasoning system. Mark Chen described it as a “risen paradigm” where models “pause to think before delivering more intelligent responses.” This eliminates the need for users to choose between a fast model and a more thoughtful one. GPT-5 automatically determines the “perfect amount” of thinking required for a given prompt, delivering deeper, more accurate answers for complex tasks while remaining fast for simpler queries. This “thinking” process is transparent; users can expand a “train of thought” to see the model’s reasoning steps, such as planning, tool selection, and fact-checking.

Performance and Benchmarks

Max Schwarzer presented data showing GPT-5 setting new state-of-the-art records across a wide range of academic and real-world evaluations, significantly outperforming previous OpenAI models and competitors.

• Coding: GPT-5 achieves a new high score on SWEBench (74.9% vs. 69.1% for the previous model), which evaluates performance on real-world software engineering tasks. It also excels on Aider Polyglot (88%), measuring its ability across multiple programming languages.
• General Reasoning: It outperforms most human experts on MMLU (Massive Multitask Language Understanding) and shows exceptional mathematical reasoning on the AIME exam benchmark.
• Factuality: A major focus was placed on reducing hallucinations. GPT-5 is presented as OpenAI’s most factual and reliable model, with significant improvements in handling open-ended and complex questions.
• Agentic Tool Use: On the Tower-Square benchmark, which tests a model’s ability to use tools to solve complex problems, GPT-5 scores 97%, a massive leap from the previous state-of-the-art of 49%.

“Software on Demand”: The Developer Revolution

A recurring theme was GPT-5’s ability to act as an agentic software developer, a concept Sam Altman called “software on demand.” Demos showcased the model’s capacity to build entire applications from a single, high-level prompt.

• Live Coding Demos:
  • An interactive physics visualization of the Bernoulli Effect was created in under two minutes, a task that would have previously taken a human developer a week.
  • A full-featured French-learning web application—complete with flashcards, quizzes, progress tracking, and an educational snake-like game with sound effects—was built from scratch in minutes.
• Pair Programming Personality: The model is tuned to be a collaborative teammate for developers. When used within an IDE like Cursor, it exhibits key personality traits: autonomy, collaboration, communication, and context management. It explains its plan, provides updates, uses best practices, runs tests, and can even fix bugs in a codebase that previous models (and humans) could not.
• Front-End and Aesthetics: GPT-5 demonstrates a sophisticated understanding of UI/UX design, creating “beautiful, tastefully designed” dashboards and 3D games. The team trained it on principles of typography, color, and spacing, allowing it to produce aesthetically pleasing results by default.

New API and Developer Features

The GPT-5 API is designed for real-world utility, moving beyond benchmark performance.

• Models and Pricing: Three models are available: GPT-5 (at $1.25/M input tokens), GPT-5 mini, and GPT-5 nano (up to 25x cheaper), allowing developers to choose the right point on the cost-latency curve.
• New Parameters: A reasoning_effort parameter allows developers to control how much “thinking” the model does, while a verbosity parameter controls the length of the output.
• Enhanced Tool Use:
  • Custom Tools: Now support free-form plaintext arguments, making it easier to pass long strings like code blocks without complex JSON escaping.
  • Structured Outputs: Developers can constrain model output to a specific format using regular expressions or context-free grammars.
  • Tool Call Preambles: The model can be instructed to output an explanation of its plan before executing a tool call.
• Longer Context: The context window is increased to 400K tokens, with state-of-the-art performance on long-context reasoning benchmarks like OpenAI MRCR and GraphWalks BFS.

Advanced Training and Safety

• Synthetic Curriculum: Sebastien Bubeck revealed a key research breakthrough: using previous generation models to generate a “high-quality synthetic curriculum” to teach GPT-5. The focus is not on creating more data, but the right kind of data, shaped specifically “to teach rather than just to fill space.” This points to a future of “recursive set of improvement loops.”
• Safe Completions: Safety systems have been overhauled. Instead of a binary “comply or refuse” response, GPT-5 uses “Safe Completions.” For dual-use or sensitive prompts, it aims to maximize helpfulness within safety constraints. This may involve partially answering, explaining why it cannot fully comply, and guiding the conversation toward safer alternatives, reducing frustrating refusals for legitimate queries.

Availability

GPT-5 is rolling out immediately to API users. In ChatGPT, it is available for free users with usage limits (falling back to a smaller GPT-5 mini model), with significantly higher limits for Plus, Pro, Team, and Enterprise users. New features like Gmail/Calendar integration and Personalities are rolling out to paid users first.

• Tags:
• Software Development
• Agentic AI
• Synthetic Data
• Generative AI
• Large Language Models
• API
• Reasoning
• GPT-5