OnPrevious previewAfter that, OpenAI officially announced the launch ofNew large-scale language model GPT-5, emphasizing comprehensive evolution in computing speed, comprehension ability, reasoning logic, context processing and structured thinking, and touting this as the "smartest, fastest and most practical" model version, which will further accelerate the pace of enterprises introducing the application of generative AI technology.
The GPT-5 model builds on GPT-4 Turbo, integrating OpenAI's previous breakthroughs in GPT-4o and o-series models, inference agents, and mathematical computing. It also provides optimized API access performance and the ChatGPT user experience through a more unified model architecture.
OpenAI stated that companies including BNY Mellon, California State University, Figma, Intercom, Lowe's, Morgan Stanley, SoftBank and T-Mobile have already introduced AI application functions through ChatGPT Team or Enterprise solutions to rebuild their internal workflows, indicating that more companies are beginning to focus on introducing AI tool resources.
In actual application feedback, companies such as Amgen, Uber, Moderna, and Salesforce have all stated that GPT-5 performs better in tasks that rely on contextual understanding and high accuracy. For example, pharmaceutical company Amgen noted that internal testing results showed that GPT-5 achieved significantly higher accuracy and consistency than previous versions when processing highly scientific content, and was also more reliable in reasoning, judgment, and contextual understanding.
OpenAI stated that currently about 7 million people use ChatGPT every week. As consumers and corporate users become more familiar with AI technology, more organizations choose to directly equip their employees with generative AI tools, hoping to achieve a leap in decision-making, creativity and efficiency.
First open to users of ChatGPT Team plan
GPT-5 is available to users of the ChatGPT Team plan simultaneously with its launch, with Enterprise and Education plans expected to be available next week. A developer API has also officially launched, allowing businesses to create more advanced application scenarios.
In addition, OpenAI also announced the upcoming release of the GPT-5 Pro version, which will provide more advanced reasoning and answer quality, targeting commercial and scientific research applications that require deeper AI inference capabilities.
With the launch of GPT-5, OpenAI aims to continue to consolidate its leadership in the field of generative AI, and sees this as a key step in making AI a core operation and innovation for future enterprises.
GPT-5 API is provided to developers and enterprises at the same time, with three model specifications
OpenAISimultaneous announcementThe GPT-5 model, which is accessible through an API, will be divided into three model specifications: gpt-5, gpt-5-mini, and gpt-5-nano, allowing developers to choose based on performance, latency, and cost flexibility.
According to data released by OpenAI, GPT-5 achieved 74.9% on the SWE-bench Verified test and 88% on the Aider polyglot test, achieving the best performance among current models of its class. Furthermore, GPT-5 can not only generate code but also fix bugs, understand large-scale program architectures, and even provide explanations before and after use, demonstrating high operability and collaboration capabilities.
In front-end development, GPT-5 outperformed OpenAI's previously released o70 model 3% of the time in internal testing and has received high praise from numerous developers. For example, the programming platform Cursor declared GPT-5 the "smartest model" it has ever used, saying it is not only easy to use but also possesses a "personality" not seen in other models. Windsurf, a developer tool provider, noted that GPT-5's tool call error rate is only half that of other leading models. Vercel even described GPT-5 as "currently the best front-end AI model," achieving top-tier standards in both aesthetics and code quality.
Beyond programming, GPT-5 also excels at long-term, multi-stage intelligent agent tasks. In the τ2-bench telecom tool call test, GPT-5 achieved a score of 96.7%.
OpenAI claims that GPT-5 performs more stably in sequential or parallel multi-tool chaining tasks, accurately executing complex tasks and effectively handling tool errors and long text context. Enterprise user Manus declared GPT-5 the "best-performing single model" in its internal evaluations, while the note-taking platform Notion emphasizes its low latency and fast response, making it ideal for scenarios requiring the completion of complex tasks in a single pass.
To give developers more granular control over model output, OpenAI has added several new features to GPT-5, such as the ability to adjust the length of responses, control the depth of reasoning, and control the speed of responses. Furthermore, GPT-5 has added custom tool functionality, allowing tools to be called using text format (rather than JSON) and applying context-free grammar constraints defined by the developer.
OpenAI also explained that the GPT-5 currently available in the API is a fully inferential version, distinct from the multi-model system used in ChatGPT. Developers can expect faster response times using the minimal inference mode, a specialized version tuned for API development needs and distinct from the non-inferential model in ChatGPT.
The launch of GPT-5 not only provides developers with more powerful tools but also further pushes the boundaries of AI performance in real-world applications. Whether it's programming, tool integration, or cross-stage task automation, this new model enables greater efficiency and accuracy.
API access pricing and deep collaboration with Microsoft
Pricing for the GPT-5 model is $1.25 per million tokens of input and $10 per output. The streamlined version, gpt-5-mini, is priced at $0.25 per million tokens of input and $2 per million output. The smallest model, gpt-5-nano, is just $0.05 and $0.4, respectively. All three models support the Responses API and Chat Completions API, and are pre-integrated with the Codex CLI platform.
In addition to the full version, OpenAI also offers a non-inference version, gpt-5-chat-latest, for general ChatGPT applications. As a lightweight option, gpt-5-chat-latest is priced the same as the standard GPT-5 version, allowing companies to choose the most appropriate model for their product integration, balancing performance and cost control.
Currently, the GPT-5 model has been fully integrated into the Microsoft ecosystem, including Microsoft 365 Copilot, Copilot, GitHub Copilot, and Azure AI Foundry platforms, further strengthening the strategic partnership between OpenAI and Microsoft.
