OpenAI on Thursday unveiled its highly anticipated GPT-5, a powerful multi-modal AI model featuring major advancements in problem-solving and coding.
The new flagship model was announced during a Thursday morning livestream. Unlike previous releases that were limited to paid subscribers, GPT-5 will be available to free-tier ChatGPT users as well, OpenAI said.
OpenAI CEO Sam Altman described interacting with the new model as akin to conversing with a PhD-level expert, noting that while GPT-3 was comparable to a high school student and GPT-4 to a college student, “With GPT-5 you get an entire team of PhD experts in your pocket, ready to help you,” Altman said.
During the announcement, OpenAI researchers emphasized that GPT-5 was designed to be more reliable and accurate, with fewer hallucinations. The company says the new model represents a significant performance improvement in three of ChatGPT’s most common uses: writing, coding, and health queries.
GPT-5 offers improvements in reasoning, higher-quality code generation, greater autonomy with reduced need for user input, and seamless integration with platforms like ChatGPT and Google’s Gmail and Calendar apps.
Microsoft said in a separate announcement that GPT-5 is now available across Microsoft platforms starting today, including Copilot, Microsoft 365 Copilot, GitHub Copilot, Visual Studio Code, and more. Developers can use GPT-5 in Visual Studio Code to write, test and deploy code with GitHub Copilot, and develop agents using the Azure AI Foundry extension.
Earlier language models relied only on pretraining to generate responses. GPT-5, like recent inference-based models, can also incorporate new data from user prompts in real time (a method known as “test-time computing”). This release also unifies OpenAI’s model naming, replacing names like o1 and o4-mini with the GPT-5 family, signaling a shift to models that combine both pretraining and inference.
For many, the key question is whether the leap from GPT‑4 to GPT-5 will prove as dramatic as the jump from GPT‑3 to GPT‑4. In OpenAI’s own testing of GPT‑5 against commonly used benchmarks, the new system appears to show significant performance increases in key areas over its previous models. These areas include coding, reasoning, instructive tasks, long-context comprehension (used during “inference” reasoning), and factual accuracy.
In coding, the system scored 74.9% on the SWE-bench Verified benchmark, up from the 69.1% earned by its “o3” reasoning model. In factuality and hallucination reduction, the system dropped to a 0.7% error rate in the LongFact-Concepts benchmark, down from the o3 model’s 4.5%, and a 1% error rate in the FactScore benchmark compared to o3’s 5.7%. In agentic task handling and tool interaction, GPT-5 (with reasoning) scored 69.6% on the Scale MultiChallenge (multi-turn instruction following) benchmark, a big increase over the OpenAI o3 reasoning model’s 60.4% score.
Industry analysts have long speculated that OpenAI’s general models are actually systems of models with unique strengths that work together to solve problems. OpenAI seems to confirm that idea in a blog post with its description of GPT‑5 as a “unified system with a smart, efficient model that answers most questions, a deeper reasoning model (GPT‑5 thinking) for harder problems, and a real‑time router that quickly decides which to use.” The company says, however, that it intends to build the router and the models into a single model “in the near future.”
OpenAI adds that when users reach their usage limits, a “mini” version of each model handles the remaining questions.
OpenAI also recently introduced two “open-weight” models—gpt-oss-120b and gpt-oss-20b—which are freely available and modifiable by developers. This move marked a rare shift from its typically closed model strategy. However, with GPT-5, OpenAI returns to its more traditional “closed” approach.
In addition, the company announced a partnership with the U.S. federal government to provide executive branch agencies access to its enterprise-grade chatbot. Through a landmark agreement with the General Services Administration (GSA), ChatGPT Enterprise will be made available to agencies for just $1 per agency for one year. OpenAI has assured that it will not use government data to train its AI models.
Update: This story now includes additional details about GPT-5 revealed after the initial announcement.
The early-rate deadline for Fast Company’s Most Innovative Companies Awards is Friday, September 5, at 11:59 p.m. PT. Apply today.
