OpenAI drops GPT-5 and it decides when to giga-think so you don’t have to

Key Takeaways

OpenAI has launched GPT-5, its most powerful AI model yet, bringing significant gains in reasoning, accuracy, safety, and real-world usefulness across domains.
The model is now the default in ChatGPT, with Pro subscribers getting access to GPT-5 Pro.

OpenAI has rolled out GPT-5, its latest AI model featuring enhanced capabilities across coding, math, writing, health, and visual perception tasks.

The company promotes the new system as highly capable and user-friendly, featuring built-in reasoning that enables it to tackle complex tasks and deliver expert-level intelligence to all users.

The unified system also comes with a smart router that automatically selects the appropriate mode based on context. The router adapts in real-time to user behavior, aiming to deliver the right level of reasoning without added latency.

“First, GPT-5 in an integrated model, meaning no more model switcher and it decides when it needs to think harder or not,” said OpenAI CEO Sam Altman in a statement. “It is very smart, intuitive, and fast.”

GPT-5 demonstrates improved performance in several key areas. According to OpenAI, the new model sets new state-of-the-art (SOTA) results across benchmarks in math (94.6% on AIME 2025), real-world coding (74.9% on SWE-bench Verified), and science (88.4% on GPQA).

GPT-5 Pro offers even longer, more accurate answers using efficient parallel computing.

As noted, the model shows a reduced tendency for incorrect information, with responses “45% less likely to contain a factual error than GPT-4o” when using web search. With thinking mode enabled, GPT-5’s error rate drops further, becoming “80% less likely to contain a factual error than OpenAI o3.”

OpenAI has made the system available to all users, with tiered access levels. Pro subscribers receive unlimited access and exclusive use of GPT-5 Pro, while Plus users get higher usage limits compared to free users. Enterprise, Team, and Education access will begin rolling out in a week.

The model incorporates new safety measures, including “safe completions” training that aims to provide helpful answers while maintaining security boundaries. OpenAI reports conducting 5,000 hours of safety testing with partners including CAISI and UK AISI.