OpenAI announces 'GPT-5,' a new model that combines the standard model and the inference model, available to all users, including those on the ChatGPT free plan
OpenAI has announced its new flagship model, GPT-5 . GPT-5 is said to be 'smarter, faster, and less likely to give inaccurate answers.' Three models, GPT-5, GPT-5 mini, and GPT-5 nano, will be available to all ChatGPT users, including free users. Paid Pro users can also access GPT-5 Pro, which uses augmented inference to provide even more comprehensive and accurate answers.
GPT-5 is here | OpenAI
Introducing GPT-5 - YouTube
Introducing GPT‑5 for developers | OpenAI
https://openai.com/index/introducing-gpt-5-for-developers/
OpenAI describes GPT-5 as the 'most advanced AI system,' claiming it represents a significant leap forward from previous models and demonstrates cutting-edge performance in a wide range of fields, including coding, mathematics, writing, healthcare, and visual recognition.
Mark Chen, chief officer of OpenAI, said that GPT-5 has introduced the concept of 'pause to think,' which allows the AI to generate deeper and more accurate responses by inserting an internal thought process before returning the output. This is the opposite of simply speeding up the response speed, and goes beyond the conventional idea of 'faster is better,' and is a new approach to achieving both quality and depth.
To achieve this, GPT-5 consists of a standard model that can answer a wide range of questions quickly and efficiently, an inference model (GPT-5 thinking) that can handle more complex problems, and a real-time router function that automatically selects the optimal model.
The router function learns user data such as the timing of model switching, satisfaction with answers, and accuracy, and improves accuracy. OpenAI plans to integrate these functions into a single model in the future.
These improvements in conversational accuracy have led to significant progress in reducing
The hallucination rate of GPT-5 was reduced by approximately 20% compared to GPT-4o, and by approximately 70% compared to the OpenAI o3 model when using the thinking function. In particular, in the benchmark for factuality of long texts (Long Fact, FActScore), the hallucination rate was reduced to approximately one-sixth of that of the o3 model.
Below is a graph comparing the hallucination rate for open source prompts between GPT-5 and the o3 model.
According to OpenAI, GPT-5 has made great strides from previous models in a wide range of fields, including coding, mathematics, and medicine, and has achieved new state-of-the-art results in many benchmarks.
GPT-5's performance improvement is particularly notable in specialized fields. In mathematics, it achieved a high accuracy rate of 94.6% on the AIME 2025 math test for American high school students, even without using any tools.
In practical software evaluation benchmarks, it achieved a score of 74.9% on SWE-bench and 88% on Aider-Polyglot. Specific capabilities include significant advances in generating complex front-ends and debugging large repositories, making it possible to create intuitive and well-designed websites, apps, games, and more with just a single prompt.
In addition, the thinking function 'GPT-5 thinking' has demonstrated superior performance in agent-based coding tasks while reducing output tokens by 50-80% compared to OpenAI o3. Furthermore, the higher-level model, GPT-5 Pro, has been reported to show particularly excellent results in the coding field.
OpenAI will be available immediately to all ChatGPT users. ChatGPT adds four personality themes—'Sarcastic,' 'Robot,' 'Listener,' and 'Nerd'—to customize how users respond. It also allows users to change the color of individual chat threads.
However, there is a limit to the number of GPT-5 users can use for free, and once that limit is reached, the model will be switched to the less powerful GPT-5 mini. Paid Plus users have a higher limit on the number of times they can use GPT-5. Furthermore, paid Pro users have access to GPT-5 Pro, which uses augmented inference to provide more comprehensive and accurate answers.
For developers who access GPT-5 via OpenAI's API, three models are available: GPT-5, GPT-5 mini, and GPT-5 nano. The pricing for each API is as follows:
'I kind of hate the term AGI because everyone uses it a little differently right now. But GPT-5 is a big step toward truly capable models. It's clearly a model with general intelligence,' Altman said. 'GPT-5 is still missing something very important. It's not a model that continuously learns from new discoveries as it's deployed. I feel like GPT-5 will definitely be part of AGI.'
Related Posts:
in Mobile, Software, Web Service, Video, Posted by log1i_yk