I Pushed GPT-6 Astra Computer Use to Its Limits

Direct Answer

GPT-6 Astra computer use is already useful for delegated visual work, but it is not a universal replacement for skilled operators or direct software integrations. Futurepedia's ten-app test found three good fits: completing a task inside unfamiliar software, handling repetitive setup inside a familiar tool, and producing an editable first pass while the human works elsewhere.

The failures are just as instructive. Forced mouse-and-keyboard control performed poorly on a complex Illustrator recreation and took almost twice as long as Python in Blender. Computer use is the fallback for software that exposes no better interface, not the default merely because watching an agent move a cursor looks impressive.

Watch the Ten-App Test

Credit and evidence note: application results, timings, prompts, and judgments come from Futurepedia's creator test, published 15 September 2026. The video deliberately blocks scripts, APIs, and MCP tools in several trials to isolate cursor-based computer use. That makes the comparison useful, but unlike a normal production workflow.

The Ten-App Scorecard

These are demonstrations by one creator, not standardized benchmarks. The creator accelerated the footage, used different prompt depths, and sometimes had prior familiarity with the software. The useful evidence is the pattern across tasks, not a universal ranking.

Creative Software: Good Assistant, Uneven Art Director

Google Earth Studio produced the clearest end-to-end win. The first pass took 11 minutes 49 seconds and established the route, camera, and keyframes. Three follow-ups added roughly 23 minutes and corrected pacing, direction, and the approach to Delicate Arch. For someone who uses the software infrequently, a usable result in about 35 minutes can be valuable even when the agent is not faster than an expert.

Canva revealed a different limitation. The agent spent under four minutes on the first drawing, 7 minutes 14 seconds on the second, and 20 minutes 47 seconds on the most detailed version. Complexity increased, but each concept repeated a similar animal-carrying-a-structure motif. Longer work did not automatically produce broader creative exploration.

Photoshop took roughly an hour across three rounds. The final file retained editable layers, which matters more than a flattened preview, but the creator could have assembled it faster himself. The gain was asynchronous labor: he reviewed results from his phone while the agent worked on a separate monitor.

Premiere was the most practical Adobe result. Astra added text behind the subject, animated placed images, searched a sound-effects project, and synchronized the choices. Twelve minutes for the first pass plus 13 minutes for the revision is reasonable for a bounded edit that would otherwise require manual keyframing.

Illustrator exposed the ceiling. The forced interface run spent 40 minutes on a poor first pass and another 52 minutes without approaching the reference. When allowed to use Python, Astra made a better version in 20 minutes and improved it over two further passes. Neither matched the target, but the direct method was plainly more effective.

LEGO, Rockets, and Physics: Completion Needs Validation

BrickLink Studio was a striking first try: Astra represented Delicate Arch with 257 pieces and organized the model into 27 assembly steps. That demonstrates competent manipulation of an unfamiliar structured design tool. It does not prove that every connection is physically robust or that the final parts list is economical.

In OpenRocket, the agent iterated until its simulated design exceeded 1,000 meters and met the recovery constraint. That is a successful software workflow, not an engineering certification. Simulation inputs, component availability, stability margins, launch conditions, and applicable safety rules still need review by someone qualified.

The Algodoo chain reaction completed the requested sequence but did not look especially difficult. This is a useful reminder to set a quality bar before starting. “The simulation runs” and “the design is elegant, robust, and worth keeping” are different acceptance tests.

Blender Proves Why the Control Surface Matters

The cursor-only Blender attempt began with a weak result after 15 minutes. Astra researched reference images, rebuilt the form, adjusted materials through the node editor, inspected screenshots, and improved the surrounding scene. After roughly one hour 45 minutes, it reached a respectable but still intermediate-looking result.

The unrestricted run used Blender's Python interface and produced a visibly better scene in about 54 minutes. The comparison does not show that computer use failed. It shows that visual understanding and direct structured control work best together. Cursor control is valuable for surfaces that have no reliable API; it should not replace a better interface to make the automation feel more human.

GarageBand Shows Adaptation, Not Musicianship

Using iPhone Mirroring, Astra tapped virtual drums and instruments rather than arranging notes directly in a desktop timeline. It could not hear the metronome in real time, press multiple points simultaneously, or reliably hold a note. It compensated by using quantization, splitting drums across tracks, and toggling sustain to extend notes.

That adaptation is the interesting capability. The final seven-track song took 52 minutes and was musically basic, but the agent identified input constraints and changed strategy without being explicitly told how. A future workflow should specify genre, tempo, reference tracks, instrumentation, song structure, and an evaluation checklist rather than simply asking for something “more complex.”

Choose the Best Authorized Control Surface

OpenAI describes Astra as its strongest computer-use model and reports better performance and lower simulated task time than GPT-5.6 Sol on OSWorld 2.0. Those measurements support the direction of improvement; they do not guarantee success in a specific application. Evaluate on the software, files, and acceptance criteria that matter to you.

A Safe First Delegation

1. Duplicate the source file. Work in a copy of the Premiere project, Photoshop document, Blender scene, or other artifact.
2. Limit access. Expose only the folders, applications, sites, and accounts needed for the task.
3. Define “done.” Specify the deliverable, editable format, visual references, constraints, and quality checks.
4. Choose the method deliberately. Permit the safest effective API or script when appropriate instead of requiring cursor control.
5. Cap the loop. Set a maximum number of attempts, time limit, or usage budget.
6. Gate consequences. Require approval before publishing, sending, purchasing, overwriting originals, deleting, or changing permissions.
7. Inspect the artifact. Review layers, timelines, geometry, formulas, simulation inputs, and exported output, not only a screenshot.

OpenAI's desktop guidance distinguishes the built-in browser from a connected Chrome profile: the former uses its own browser state, while the Chrome route is for existing signed-in sessions and extensions. That difference should influence which accounts and data you expose.

Video Chapters

Verdict

Astra's computer use is most valuable as asynchronous, cross-application labor. It can get a non-expert surprisingly far inside an unfamiliar tool and can remove setup or repetitive work from an expert's workflow. It remains slower than a skilled operator on some tasks, and the visual interface can be a severe bottleneck when the software provides a structured control path.

The right adoption question is not “Can it use this app?” It is “Can it return an editable, reviewable result at a lower cost in human attention, inside boundaries we trust?” This test provides several credible yeses, several clear noes, and a much better way to decide.

Sources and Further Reading

• Futurepedia: I Pushed GPT-6 Astra's Computer Use to Its Limits (creator test, 15 September 2026)
• OpenAI: GPT-6 Astra (official capability, benchmark, availability, and safety overview)
• OpenAI Help: ChatGPT Work and Codex and built-in browser guidance

Publication date follows the primary video's official YouTube date: 15 September 2026. Editorial review: 17 September 2026. Product access, pricing, usage limits, and interface behavior may change.