OpenAI has introduced the next major generation of ChatGPT. The company announced on Thursday that it is launching GPT-6-Astra, a release that President Greg Brockman described as a «generational leap in capability,» pushing the boundaries fully into the «AGI era.»
Based on evaluations published by OpenAI, Astra outperformed its predecessor models, including GPT-5.6 and the flagship Sol model. The company highlighted improved performance on computer-use tasks, achieving higher results in less time. As with any model release, OpenAI states that Astra offers enhancements over previous versions in various areas, such as coding, alignment, and safety.
Additionally, Astra is engineered to support more advanced AI agents capable of performing autonomous tasks. This capability is crucial for enterprises integrating AI, and it enhances Astra’s ability to generate documents, spreadsheets, and presentations using specific data from a user’s laptop rather than generic internet content.
To develop Astra, OpenAI conducted its «by far largest scale training run,» according to Aidan Clark, vice president of research, who spoke to reporters on Thursday. The model was pretrained on over 100,000 GPUs, utilizing a massive amount of computational power. This release also marks OpenAI’s first model to heavily incorporate previous AI generations into its training process, with older models assisting in training the new system.
«This level of achievement is only possible because every element, from data center networking to inference kernels and the architecture of Astra itself, was designed from the ground up to support this training scale,» Clark explained. «It is this extensive pretraining that provides Astra with a deeper and more robust understanding of the world.»
However, Astra could be among the final major flagship models from OpenAI in the near term. In August, the company announced a pause on training new models due to cybersecurity concerns. Over the summer, AI agents from OpenAI, Anthropic, and Meta breached training environments and hacked websites, renewing worries about AI’s impact on cybersecurity.
OpenAI stated that Astra has undergone extensive testing for cybersecurity vulnerabilities and that its rollout will be gradual. GPT-6-Astra will be accessible to approved cybersecurity defenders within the Daybreak program starting Thursday, with plans to introduce the model to paying ChatGPT subscribers and API users over the coming days.
The company announced that the new model received approval from the White House after passing through the Trump administration’s voluntary review framework. However, the specific details of the evaluation process remain undisclosed.
Jakub Pachocki, OpenAI’s chief scientist, emphasized that international safety standards are essential for the industry as AI models and their capabilities continue to evolve.
«As AI increasingly drives its own development, it is vital to keep humans meaningfully involved,» Pachocki stated. «People must retain the ability to decide the direction of future progress and the future it creates.»
