GPT-6 Astra Adds Asynchronous Tool Calls and In-Flight Steering
OpenAI developer experience engineer Charlie Guo presents GPT-6 Astra as a model for longer, less-defined assignments that require it to retain context across tools, applications and changing user instructions. His central case is that asynchronous tool calling and in-flight steering in the Responses API let developers keep work moving—and redirect it—without cancelling a slow operation or rebuilding the task, alongside improved computer use and stronger creative, knowledge-work and 3D outputs.
Async tools and steering keep work moving without restarting it
? charlie-guo says GPT-6 Astra adds two related capabilities to the Responses API: asynchronous tool calling and steering. Together, they are intended to let a model continue work while a tool call is still running, then accept new instructions before an active response is complete.
OpenAI’s displayed configuration marks a render_scene function as asynchronous with async: true. Guo says that lets the model work on other parts of an assignment while the call runs, then incorporate its result when it becomes available. For a longer task, the practical implication is that a slow tool call need not stop all progress.
Steering provides a way to alter that task in flight. The implementation OpenAI shows sends a WebSocket message of type response.steer, identifies the previous_response_id, and supplies new input. Guo compares the interaction to steering in Codex: rather than waiting for a response to finish, a user can add context or redirect the model while it is working.
The Japanese-garden demonstration combines both features. Astra is asked to build and refine a scene with an open pond, an arched bridge, and a pavilion framed by trees. While render_scene is active, the user changes the request: “Actually, let’s make the trees red.” The interface shows the instruction accepted as response.steer.accepted while the rendering continues.
I can change the direction without having to cancel the running tool or start the task over, and I get the final result with the new context incorporated.
Guo’s argument is that extended work needs an intervention mechanism that does not require reconstructing the assignment from scratch. The controls matter most where both conditions hold: external tools take time, and the user’s requirements may change before those tools return.
Computer use is presented as persistent visual context across an application
? charlie-guo describes computer use as a model using a computer or browser as a person would. He says Astra is more accurate and efficient at the capability, and that the improved computer-use skills are available in both the ChatGPT app and the API.
His demonstration begins with a photograph and a request for Codex to use Krita to paint it in the style of Van Gogh, adding the Golden Gate Bridge in the background. The task asks the model to retain the image and creative constraints while navigating a digital-painting application.
“Astra has to keep track of the picture I asked for while working through the app’s controls,” Guo says. The Krita playback, shown at 28× speed, produces a painting of a laptop on a desk, with the Golden Gate Bridge visible through a window.
Guo says Astra uses screenshots to observe what is happening in the application while keeping the app in the background, allowing the user to focus elsewhere. For developers, his examples are filling out forms for an API key and performing QA on a mobile app. The relevant capability, as he frames it, is not just selecting controls on a screen; it is carrying a visual brief and task context through a sequence of application actions.
OpenAI emphasizes detail, prompt understanding, and 3D work
? charlie-guo presents Astra’s creative and knowledge-work performance through side-by-side outputs rebuilt from prompts used for earlier demos, games, and tools. OpenAI displays GPT-5.5, GPT-5.6 Sol, and GPT-6 Astra across projects including a waveform studio, a watchmaker landing page, a Codex pet arena, and a Golden Gate experience.
Guo says Astra had more attention to detail, a better understanding of the user’s prompt, and more sophisticated outputs across those examples. He identifies 3D work as a particular strength. The shown montage includes a Japanese garden over water, a spaceship fleet-registry interface, a pelican riding a bicycle, a cityscape, and a Dyson sphere.
Guo makes a separate, personal assessment of Astra’s text output and collaborative feel. In his own use, he says, its slideshows and essays have been significantly easier to read and work with than outputs from other models he has tried. He also says its personality feels more natural, particularly during back-and-forth feedback.
He frames those claims as part of Astra’s usefulness on work that is larger and less well defined. Guo says it feels like working with an experienced collaborator: a model he can give bigger assignments with minimal hand-holding, across coding, writing, design, and other work requiring what he calls more intellectual horsepower. He says it often delivers strong results on the first attempt.
According to Guo, Astra is available in ChatGPT, Codex, and the API.