OpenAI: Astra Can Handle All Operations on Computers
OpenAI's official account defined GPT-6 Astra: Anything that can be done on a computer, Astra can handle, and faster. The official website positions it as the strongest model for computer usage, browsing, software engineering, cybersecurity, and scientific and professional work.
The company claims Astra scored 72.6% in computer usage during the OSWorld 2.0 delayed simulation, with a single task taking about 40 minutes, compared to GPT-5.6 Sol's 65.7% and about 75 minutes, saving approximately 47% time. Examples include filling out web forms, modifying customer relationship systems, organizing calendars, retrieving and writing emails or documents, analyzing scientific data and generating graphs, building websites and performing frontend acceptance, and troubleshooting software installation based on screen failures. The interface documentation lists computer usage, hosting shells, patches, skills, model context protocols, and asynchronous tool calls: the model can continue reasoning while executing tools in applications. The pricing is $10 for input and $50 for output per million tokens, with tools charged separately. Initially open to enterprises in the trusted access program, followed by Plus to enterprise subscriptions and interfaces with Amazon Cloud. ExploitBench's full score and advanced network capabilities still go through restricted channels. The enterprise workspace is closed by default.
Who is buying and who is selling: The event is driven by selling chat windows as remote employees, not by the launch of a new operating system. Buyers are knowledge workers looking to outsource form filling, acceptance, and troubleshooting, while sellers are laboratories charging by tokens and tool usage. Beneficiaries include the Codex path integrated into ChatGPT, while those under pressure are software vendors selling remote desktops and process robots by seat. Funds flow from subscriptions and tool fees into computer usage reasoning, with speed advantages measured in minutes and risk from permissions hedged by the default closure for enterprises.
Source: Public Information
ABAB AI Insight
"Anything that can be done on a computer can be handled" transforms the product from a Q&A machine into an operator. The scoring shifts from accuracy on questions to time spent in OSWorld, indicating that the competitive dimension is now whether one can continuously complete a set of professional software. Asynchronous tool calls are designed to keep the model running while waiting for interfaces, making the agent a overlapping assembly line.
The capital path is to exchange high output unit prices for less total time: token counts per task decrease, but bills may not necessarily rise. Money flows into the computer usage toolchain and hosting shells, rather than piling up another round of pure dialogue. Trusted access is initially provided to enterprises, separating the demonstration environment from production permissions.
Comparative objects include Claude's computer usage, the baseline of 75 minutes from its own GPT-5.6 Sol, and traditional RPA based on scripted clicks. The industry position shifts from assisting writing to expanding operational automation, with control over screen permissions and tool billing, rather than parameter slogans.
Structural change is technological substitution. The mechanism is that once keyboard and mouse processes can be stably completed by the model, the role shifts from "people who can operate software" to "people who dare to hand over the screen to the model"; substitution occurs in minute-level tasks, while pricing occurs on tool calls.
ABAB News · Cognitive Laws
Being able to answer questions does not equal being able to complete software clicks
Tasks become shorter, and bills are recalculated based on tool usage
Permissions that are closed by default are the true product boundaries