Today, OpenAI released GPT-6 Astra, marking a significant advancement in its AI capabilities. The model is positioned primarily as a computer-use system rather than a chat model, designed to operate software like a human across various platforms including browsers, spreadsheets, desktop applications, and terminals. Astra completes multi-step tasks autonomously instead of merely providing instructions.
Astra is a closed, hosted model available only to organizations within OpenAI’s Trusted Access and Daybreak programs. It is not self-hostable as OpenAI has not released the model weights. The model features a 1,050,000-token context window and supports text and image input, producing text-only output. Astra introduces new reasoning levels, xhigh and max, and supports several tools for computer use, although fine-tuning is not available. Pricing is set at $10 per million input tokens and $50 per million output tokens, with adjustments for larger requests.
Astra's technical improvements include enhanced context handling, which retains detailed information across context windows, allowing it to refer back to earlier messages and tool outputs. This approach addresses limitations of previous models, where context compaction led to the loss of important details. Additionally, Astra can continue tasks independently while awaiting user input, mitigating common workflow stalls. The model demonstrated superior performance in benchmarks, scoring 72.6% on OSWorld V2-Offline and 99.9% on ARC-AGI-3.
The model is intended for developers and organizations looking to leverage advanced AI for software operation and cybersecurity tasks. It is particularly suited for environments requiring automation of complex, multi-step processes.
Work implications: GPT-6 Astra could significantly enhance roles involving software operation and cybersecurity, offering tools for automation and advanced task management.
Originally reported by Marktechpost Media