San Francisco, August 7, 2025: OpenAI on Wednesday announced the release of GPT-5, the latest version of its flagship generative AI model, which the company says brings major improvements in speed, reasoning, factual accuracy, and real-world usability.
The launch event, held at OpenAI’s San Francisco headquarters and livestreamed globally, showcased GPT-5 as a system designed to power a wide range of use cases, from software development and business automation to health guidance and personalized digital assistance.
“GPT-5 is faster, more accurate, and more useful across more domains,” said OpenAI CEO Sam Altman. “It’s the most capable model we’ve released, and it brings us closer to safe, everyday AI integration.”
Unified System With Dynamic Routing
Unlike previous models, GPT-5 is built as a unified system composed of multiple internal models. According to OpenAI, a new smart router dynamically decides whether to use a “fast” model for lightweight queries or a “deep reasoning” model for more complex tasks. Miniature fallback models are used to maintain responsiveness during high-traffic periods.
This architecture, now fully implemented across ChatGPT, eliminates the need for users to manually select models. OpenAI confirmed that GPT-5 replaces GPT-4, GPT-4o, o3, and o4-mini, all of which have been deprecated.
Improvements in Performance and Reliability
OpenAI claims GPT-5 significantly outperforms its predecessors in benchmarks for coding, health-related advice, factual accuracy, and reduced hallucinations.
-
Hallucination reduction: Internal evaluations show a 30% to 80% decrease in hallucinated content across tasks compared to GPT-4.
-
HealthBench performance: GPT-5 ranks highest to date on HealthBench, offering safer and more nuanced responses to health-related queries.
-
Coding tasks: The model performs better on SWE-bench and code generation tasks, with support for inline annotation and step-by-step reasoning.
The model also includes improvements in tone, personalization, and clarity. It now admits uncertainty more frequently and avoids defaulting to excessive politeness or flattery — two common criticisms of earlier versions.
Developer Access and API Features
Developers now have access to three versions of GPT-5 via the OpenAI API:
gpt-5 (full model)
gpt-5-mini (faster, lightweight version)
gpt-5-nano (optimized for edge computing)
New API features include:
Customizable reasoning_effort and verbosity parameters
Tool use via plaintext (no JSON required)
Improved integration with multimodal inputs (text, image, code)
OpenAI says these changes aim to make GPT-5 “more flexible and programmable out of the box.”
Business Adoption
Several enterprises are already deploying GPT-5 in production. OpenAI cited examples including:
Figma, using GPT-5 for automated UI prototyping
Amgen, applying the model to drug development research
Morgan Stanley, integrating it into wealth advisory systems
BNY Mellon, automating internal reporting and analysis
GPT-5 is now available to all ChatGPT users, with advanced access for Plus, Pro, and Enterprise subscribers. Free-tier users are routed to GPT-5 Mini.
Presentation Highlights
During the launch presentation, CTO Mira Murati and Chief Scientist Ilya Sutskever demonstrated GPT-5's ability to generate clean web interfaces, translate and summarize documents, and provide real-time reasoning with context. Live examples included debugging a JavaScript function, interpreting a business report, and scheduling tasks via Gmail and Calendar integrations.
The event also emphasized GPT-5’s improved behavior alignment. OpenAI says the model is designed to avoid deceptive responses and reduce “false confidence” in high-stakes topics such as healthcare and law.
Not AGI, But a Step Closer
Despite the improvements, OpenAI clarified that GPT-5 is not artificial general intelligence (AGI). However, the company described it as a “meaningful advancement toward more general-purpose AI systems.”
“GPT-5 does not reason like a human,” said Sutskever, “but it performs well enough in many reasoning tasks that it can be trusted as a productive collaborator.”
OpenAI also reiterated its commitment to safety and transparency, noting that GPT-5 has undergone more extensive red-teaming, training oversight, and behavioral tuning than any of its previous models.
Availability and Outlook
GPT-5 is live on ChatGPT as of August 7 and is being gradually rolled out across OpenAI’s API, Azure OpenAI Service, and enterprise partnerships. It supports text, code, image interpretation, and tool use.
The model will be updated continuously, and OpenAI says further enhancements to memory, multimodality, and long-context understanding are already underway.