OpenAI has unveiled GPT-5, a new AI system that unifies previous model lines into a single adaptive architecture capable of adjusting its "thinking effort" based on task complexity. The system utilizes a real-time router to switch between the fast gpt-5-main model for standard queries and the deeper gpt-5-thinking model for complex problems.
- GPT-5 achieves 74.9 percent on SWE-bench Verified and 88 percent on Aider Polyglot, reducing coding error rates by two-thirds compared to o3.
- The model scores 46.2 percent on HealthBench Hard and 94.6 percent on AIME 2025, while producing about six times fewer hallucinations than o3 on fact-based benchmarks.
- A new "Safe Completions" system replaces hard refusals, aiming to provide helpful outputs within safety guidelines after extensive red teaming.
- The API introduces three model sizes (gpt-5, gpt-5-mini, gpt-5-nano) with a context window of 272,000 input tokens and 128,000 output tokens.
- Access is tiered, with free users getting GPT-5 and GPT-5-mini, while Pro subscribers gain access to the more capable GPT-5 Pro.
OpenAI states that these improvements aim to deliver more reliable and accurate responses across programming, healthcare, and writing tasks. The rollout begins immediately for ChatGPT users and API developers.