OpenAI launches GPT-5 with integrated advanced reasoning capabilities

OpenAI has announced the launch of GPT-5, its most advanced artificial intelligence model to date, which introduces a unified system capable of automatically alternating between quick responses and deep reasoning based on the complexity of each query. The model incorporates an intelligent router that determines when to apply extended thinking capabilities, based on conversation type, complexity, and specific user needs.

Among the most notable improvements is a 45% reduction in incorrect information compared to GPT-4o when using web search, reaching up to 80% fewer inaccuracies than OpenAI's o3 model when employing reasoning capabilities. The system also shows significant advances in reducing excessively compliant behaviors, decreasing flattering responses from 14.5% to 6%.

In terms of technical performance, GPT-5 achieves scores of 94.6% in advanced mathematics (AIME 2025), 74.9% in real-world programming (SWE-bench Verified), and 84.2% in multimodal understanding (MMMU). The model demonstrates particular strength in programming, especially in frontend development and debugging large repositories, as well as creating websites and applications with a single command.

The company has implemented a new security approach called "safe responses," which allows the model to provide useful information while maintaining appropriate security boundaries. This system is particularly relevant for dual-use queries, where information can have both benign and potentially harmful applications.

OpenAI also introduces four preset personalities for ChatGPT: Cynic, Robot, Listener, and Nerd, which allow adjusting interaction style without the need for complex custom instructions. These options aim to make the experience feel less artificial and more like conversing with a specialized assistant.

In addition to the main improvements, GPT-5 incorporates specific advances in creative writing, including better handling of complex structures like free verse and unrhymed iambic pentameter. In the healthcare field, the model achieves superior scores on HealthBench, functioning as a more proactive collaborator that identifies potential medical concerns and formulates relevant questions. The system also demonstrates competence in economically relevant tasks, matching or surpassing human experts in approximately half of the evaluated cases in more than 40 professions, including law, logistics, and engineering. Regarding specific risks, OpenAI has implemented reinforced security measures to prevent misuse of the model in biological and chemical fields, after conducting 5,000 hours of specialized testing. The system also demonstrates greater transparency, reducing occasions when it provides false responses about tasks it cannot complete, from 4.8% to 2.1% compared to previous models.

The model will be available to all users, with differences in usage limits depending on subscription type. Free users will access GPT-5 with volume restrictions, while Pro subscribers will have unlimited access and access to GPT-5 Pro, a variant that uses extended compute time for particularly complex tasks. Implementation began for Plus, Pro, and Team users, with availability for Enterprise and educational accounts scheduled for next week.

With GPT-5, OpenAI seeks not only to offer a more powerful and secure model, but also to bring advanced reasoning capabilities to a much broader audience.

GPT-5 Key Points

• GPT-5 incorporates an automatic reasoning system that decides when to think more deeply based on the complexity of each query.
• The model reduces incorrect information by 45% compared to GPT-4o and up to 80% fewer inaccuracies than the o3 model when using reasoning.
• Achieves record scores of 94.6% in advanced mathematics, 74.9% in real programming, and 84.2% in multimodal understanding.
• Significantly improves in programming, especially creating complete websites and applications with a single command.
• Reduces excessively compliant responses (sycophancy) from 14.5% to 6%, making interactions more natural.
• Incorporates a new "safe responses" system that provides useful information while maintaining appropriate security boundaries.
• Demonstrates competence comparable to human experts in half of the tasks evaluated in more than 40 professions.
• Improves creative writing with better handling of complex poetic structures like free verse and iambic pentameter.
• Reduces false responses about tasks it cannot complete from 4.8% to 2.1% compared to previous models.
• Will be available to all users with different limits depending on subscription type, starting with Plus, Pro, and Team users.