OpenAI Launches GPT-5, Setting New Benchmarks in AI Performance and Reliability
OpenAI has announced the release of GPT-5, describing it as its “best AI system yet” and a major leap forward in reasoning, accuracy, and versatility. The unified model excels in coding, writing, health, and multimodal tasks, while offering significant gains in factual reliability and safety.
GPT-5 is available immediately to all ChatGPT users, with Plus subscribers enjoying higher usage limits and Pro subscribers gaining access to GPT-5 Pro, a variant optimized for extended reasoning and the most complex queries.
A Unified, Smarter System
GPT-5 integrates three components: a fast, efficient core model; GPT-5 Thinking for deeper reasoning; and a real-time router that chooses the best approach based on query complexity. Once usage limits are reached, a mini version ensures continued access.
The system has been designed to reduce hallucinations, improve instruction following, and minimize sycophancy. According to OpenAI, GPT-5 is particularly strong in three high-demand areas: writing, coding, and health.
Benchmark-Breaking Performance
In academic and real-world tests, GPT-5 sets new state-of-the-art scores, including:
- Math: 94.6% on AIME 2025 without tools
- Coding: 74.9% on SWE-bench Verified, 88% on Aider Polyglot
- Multimodal Understanding: 84.2% on MMMU
- Health: 46.2% on HealthBench Hard
For economically important tasks, GPT-5 matches or surpasses expert performance in about half of the cases across 40 professions. Its reasoning mode delivers results with 50–80% less token output than previous top models.
Advances in Key Applications
Coding: GPT-5 improves on complex front-end generation and debugging, producing responsive, visually polished applications. Early testers noted better typography, spacing, and design choices. Writing: The model handles structurally complex tasks with greater precision, supporting both creative expression and practical writing such as reports, memos, and emails.
Health: GPT-5 ranks higher than all prior models on OpenAI’s HealthBench evaluation, adapting responses to a user’s knowledge level, location, and context.
Safety, Reliability, and Customization
GPT-5 is ~45% less likely to produce factual errors than GPT-4o in real-world prompts, and up to six times less likely to hallucinate on open-ended fact-seeking questions. It also shows reduced rates of deceptive responses and integrates safe completions training for more nuanced handling of dual-use topics.
New preset personalities — Cynic, Robot, Listener, and Nerd — are available in a research preview, offering tailored interaction styles without custom prompt engineering.
GPT-5 Pro for the Most Complex Tasks
GPT-5 Pro replaces OpenAI o3-pro, delivering the best performance in the GPT-5 family. It scored highest on the GPQA benchmark and was preferred by experts in 67.8% of evaluations over GPT-5 Thinking. Its strengths span science, health, mathematics, and advanced coding.
Availability
GPT-5 is now the default ChatGPT model for signed-in users, replacing GPT-4o and other variants. It is rolling out to Plus, Pro, Team, and Free users this week, with Enterprise and Edu access coming next week.
Free users will receive GPT-5 with usage caps before switching to GPT-5 mini. Pro subscribers get unlimited GPT-5 access, including GPT-5 Pro, while Plus users receive higher limits than free accounts.
