OpenAI Releases GPT-6 Astra for Coding and Computer Use

OpenAI has unveiled GPT-6 Astra, an advanced model designed for coding, computer use, professional workflows, science, and cybersecurity. This new model is currently accessible to select organizations and will be rolled out to ChatGPT Plus, Pro, Business, and Enterprise users, along with the OpenAI API, Microsoft Azure, and AWS Bedrock.

Unlike previous models, GPT-6 Astra goes beyond mere response generation, capable of executing multi-step tasks directly within software applications. It can interact with graphical user interfaces, fill out forms, update customer relationship management records, conduct research, create websites, analyze data, install and test software, and troubleshoot issues visible on the screen.

In terms of performance, GPT-6 Astra has achieved a score of 72.6% in OSWorld 2.0, surpassing its predecessor GPT-5.6 Sol's score of 65.7%. The model demonstrates proficiency in coding with a score of 57.9% on Terminal-Bench 4.0 and 74.1% on DeepSWE v1.1.

A notable feature of GPT-6 Astra is its experimental context mechanism in Codex, which enables the model to maintain notes across context windows, unlike previous versions that relied solely on compaction. This allows the model to retain earlier requirements, test results, and tool outputs during long-running coding tasks. The model can handle long contexts of up to one million tokens in OpenAI's MRCR evaluations, scoring 96.3% in the 512K-to-1M range.

GPT-6 Astra also shows significant improvements in professional tasks such as database migrations, CAD generation, data science, browser research, and scientific workflows. In the realm of cybersecurity, GPT-6 Astra is the first OpenAI model to reach the critical cybersecurity capability level under the company's Preparedness Framework.

In unregulated testing, the model discovered and utilized two previously unknown vulnerabilities and developed exploits against hardened browsers and operating systems. However, in the production version, advanced offensive tasks are restricted, while broader defensive capabilities are provided through OpenAI's Daybreak program.

Internally, OpenAI reports a lower hallucination rate for GPT-6 Astra at 4.2%, compared to 12.2% for GPT-5.6 Sol. However, monitoring the model's written reasoning remains challenging, which is an active area of research for OpenAI.

Community reactions have been mixed, with Nvidia CEO Jensen Huang praising the infrastructure used to train the model, stating, "GPT-6 Astra, trained on ~100K+ NVIDIA Grace Blackwell NVLink72." Other reactions have echoed the notion of AGI (Artificial General Intelligence) being closer to reality with the release of ChatGPT 6 Astra.