OpenAI Unveils GPT-5: A Unified AI Leap Forward with Smarter Reasoning, Fewer Hallucinations, and Broader Access

OpenAI has launched GPT-5, its first unified AI model, marking a major milestone in the company’s journey toward artificial general intelligence. The new model, released Thursday, combines the advanced reasoning of OpenAI’s o-series with the speed and responsiveness of its GPT line, signaling a shift from chatbot-like interactions to more autonomous, agent-like behavior. CEO Sam Altman described GPT-5 as “the best model in the world” and a significant leap toward AI systems that can outperform humans in most economically valuable tasks. He called the achievement “unimaginable” just a few years ago, underscoring the rapid pace of progress in the field. Starting Thursday, GPT-5 becomes the default model for all free users of ChatGPT, a move aimed at broadening access to advanced AI reasoning capabilities. Previously, such features were reserved for paying subscribers. Nick Turley, OpenAI’s VP of ChatGPT, said the decision aligns with the company’s mission to make powerful AI tools widely available. With over 700 million weekly users, ChatGPT remains one of the most-used AI platforms globally. GPT-5’s release is being closely watched by tech leaders, investors, and regulators as a potential indicator of how fast AI is advancing—and whether it’s reaching new levels of capability. In benchmark tests, GPT-5 shows strong performance across several key areas. On SWE-bench Verified, a real-world coding evaluation, it scored 74.9% on its first try—slightly ahead of Anthropic’s Claude Opus 4.1 (74.5%) and Google DeepMind’s Gemini 2.5 Pro (59.6%). On GPQA Diamond, a PhD-level science test, GPT-5 Pro achieved 89.4%, outperforming both Claude Opus 4.1 (80.9%) and Grok 4 Heavy (88.9%). In the challenging Humanity’s Last Exam, which tests knowledge across math, science, and humanities, GPT-5 Pro scored 42% when using tools—just behind xAI’s Grok 4 Heavy at 44.4%. One of GPT-5’s standout improvements is its reduced hallucination rate. On a test measuring factual accuracy, GPT-5 (with thinking) hallucinated only 4.8% of the time—far better than GPT-4o (20.6%) and o3 (22%). In health-related queries, it made errors just 1.6% of the time, a major improvement over earlier models. OpenAI also claims GPT-5 is more trustworthy and less prone to deception. Safety lead Alex Beutel said the model is better at detecting malicious intent and rejecting harmful requests while minimizing false rejections of harmless queries—leading to a more transparent and reliable experience. For creativity and writing, Turley said GPT-5 exhibits “better taste” and more natural, engaging responses. Users can now choose from four new personalities—Cynic, Robot, Listener, and Nerd—to shape how ChatGPT responds without needing to prompt it directly. On the developer front, GPT-5 is available via API in three versions: gpt-5, gpt-5-mini, and gpt-5-nano. Developers can also control response length and verbosity. The base model costs $1.25 per million input tokens and $10 per million output tokens. Pro subscribers ($200/month) get unlimited access to GPT-5 and a more powerful GPT-5 Pro version. Team, Edu, and Enterprise users will receive access next week. The launch follows OpenAI’s release of gpt-oss, an open-weight reasoning model that matches the performance of earlier top models at a fraction of the cost. While GPT-5 sets a new benchmark in some areas—especially coding and accuracy—it performs similarly to competitors in others, like agentic task completion on Tau-bench. Ultimately, benchmarks only tell part of the story. The true test will be how developers and users apply GPT-5 in real-world scenarios, and whether it delivers on the promise of smarter, safer, and more capable AI.