Standing in front of the Golden Gate Bridge, a developer snapped a photo and typed a single instruction: paint this picture in the style of Van Gogh and add the bridge in the background. What happened next took less time than it would take most people to open a paint program. GPT-6 Astra, OpenAI’s latest frontier model, launched the app, navigated the controls, and produced a recognizable Van Gogh-style rendering autonomously. For anyone who builds with AI tools for a living, that moment signals something worth paying close attention to.
A model that keeps working while you step away
The painting task is a clean illustration of what OpenAI is calling improved computer use. Astra takes screenshots to monitor what is happening on screen, keeps the target application running in the background, and works through the interface the same way a person would. The person who sent the instruction could focus on something else entirely while the model handled the controls. The same capability extends to the kinds of tasks developers repeat daily: filling out forms to obtain an API key, running quality assurance on a mobile app, or navigating multi-step browser workflows without supervision.
The underlying architecture has also changed in ways that matter for longer jobs. OpenAI is adding asynchronous tool calling and steering to the Responses API. With async tool calling, Astra can continue working on other parts of a task while it waits for a tool call to return results, rather than sitting idle. Steering goes a step further: it lets a developer redirect the model mid-task without canceling the work already in progress. In a live demonstration, a new instruction was sent while Astra was already running, and the model picked up the updated direction, adjusted its next steps, and delivered a final result that incorporated the change without requiring a restart. The already-running tool kept going while Astra folded in the new context.
Where the output quality gap shows up most clearly
OpenAI tested the model by taking prompts from demos, games, and tools built with earlier models and rebuilding them with Astra at different reasoning levels. The results showed more attention to detail, stronger understanding of the original prompt, and more sophisticated final outputs across the board. Three-dimensional modeling stood out: the model has produced renderings of gardens, shipyards, animals, cityscapes, and Dyson spheres. Slideshows and long-form essays also came back more readable and easier to work with than comparable outputs from previous generations.
One member of the OpenAI team put the day-to-day experience plainly: ‘Astra feels like working with an experienced collaborator. I’m able to hand it bigger, less well-defined tasks with minimal handholding.’
The grounding detail that separates this release from a routine model update is the steering capability: a developer can change what the model is doing in real time, mid-response, without losing the work the model has already completed. That specific pairing of async execution and live course-correction is new to the Responses API and was not available in earlier releases.
A canvas left mid-stroke
Somewhere on a developer’s screen, a Krita canvas holds a Golden Gate Bridge rendered in thick impasto brushwork, generated without a single human hand on the stylus.
GPT-6 Astra is available now in ChatGPT, Codex, and the API.
