GPT-6 Astra Targets Long-Running Work Across Desktop Applications

OpenAI says GPT-6 Astra achieves state-of-the-art benchmark results on long-running computer-use tasks across professions and desktop applications. The company also calls it its best software-engineering model to date and its most aligned model yet, citing improvements in honesty and reductions in deceptive behavior.

The product claim is broader than answering a prompt or producing a standalone file. Astra is presented as taking spoken instructions, working through the applications where the task lives, retaining context as the user changes direction, and moving between browser pages, local files, documents, and creative software. The user describes an outcome rather than specifying menus, controls, or intermediate steps.

A short comparison puts that interaction model in a longer lineage. A 1979 title card accompanies the command, “Create a yellow circle there.” In the 2026 sequence, a user asks for a small yellow circle, turns it into the window of a rocket ship, asks for more detail, and then requests a 3D model in Blender. Astra responds, “Opening Blender,” and the Blender workspace appears with a wireframe cylinder at its center. When the user later asks for a file to send to a 3D printer, Astra says it will create an STL file of the rocket.

The consequential feature of that sequence is iterative control. The user does not need to formulate the whole design in advance: they can make an object, reinterpret it, elaborate it, move it into another tool, and request a production-ready format as the work develops.

A request for a high-end, colorful, fun presentation for retailers about the next season’s rainwear becomes a Google Slides document with a retailer-facing buy order. The visible slide, titled “Rainwear 2027,” lays out four jacket silhouettes, six drops, case sizes, size allocations, wholesale and retail values, and shipment dates. When the user asks for a background color that complements the rain jackets, the presentation is shown with Google Slides’ background-color controls open.

The same thread of work extends to selling a flea-market table. After the user asks Astra to make an eBay listing for a “wild orange table,” eBay’s homepage appears. The user then asks it to include a photograph from the Downloads folder—orange-table-dent.png—and to note that the table has a slight dent. The footage shows a file-selection dialog with that image selected, followed by eBay’s photo-editing interface and a close-up of the orange table.

The sequence makes local context part of the proposition. A usable listing needs more than text generation: it needs access to the saved image, accommodation of a newly mentioned condition issue, and work inside the marketplace’s own listing flow.

A legal-document request follows a similar pattern. A user asks for a licensing-agreement template for lawyers at their firm to review, then requests that the limitation-of-liability provision be made more favorable to the licensor. A Google Doc shows the agreement’s limitation-of-liability section, including exclusions and an aggregate-liability cap tied to amounts paid by the licensee in the prior 12 months. Astra replies that it has “tightened it so the licensor’s liability is more narrowly capped.”

The demonstration also shows a retro-styled 3D game in which a rocket navigates floating asteroids, with score and health indicators on screen. The user had asked for an asteroid-dodging game controlled by arrow keys, with the space bar used to boost. Alongside the document, marketplace, and design examples, the game situates Astra’s stated computer-use capability in both creative and interactive software work.

Astra is shown handling requests that unfold on different timelines. While the rainwear presentation and eBay listing are in progress, the user asks it to order beef and rice from a restaurant used the previous week. Astra answers, “Looking into ordering food.” The food-order outcome is not shown.

The user also asks for a tennis court in the Lower Haight that afternoon. Astra later reports finding an open court at 5 PM; the user responds, “Yeah, book it.” The footage does not show the reservation being completed.

Those exchanges sit beside work whose interfaces are shown directly: the presentation in Google Slides, the local photo selected for eBay, the legal agreement in Google Docs, the asteroid game, and the Blender workspace. Together, they present Astra as an assistant intended to hold multiple jobs in context while the user continues revising, approving, or adding to them.

OpenAI closes by directing users to download the ChatGPT desktop app for the “full Astra experience.” That instruction follows the product model shown throughout: Astra’s role depends on working across desktop software, browser-based services, documents, and files already stored on the user’s machine.

The demonstrations rely on precisely those transitions. Astra opens Blender for the rocket project, works in Google Slides on the rainwear deck, reaches into the Downloads folder for the table photo, and addresses language in a Google Doc. The value OpenAI is presenting is not simply generated content within a chat interface, but an assistant that can take an evolving brief into the tools where that content must be shaped, edited, or used.

OpenAI says Astra is rolling out immediately to enterprises and customers in its Trusted Access Program. It says broader availability in ChatGPT Work and Codex is coming to Plus, Pro, Business, and Enterprise plans, with API and AWS availability following in the next few days.