GPT-6 Astra for Real Work: Video, Browser Control, Research Apps
How GPT-6 Astra handles video editing, browser automation, and knowledge work, based on early access demos beyond the 3D game showcases.
What can GPT-6 Astra actually do for work, not just demos?
GPT-6 Astra, the newest model release covered by early access creators, has drawn most of its early attention for flashy 3D world generation and playable games built from a single prompt. But underneath the spectacle sits a smaller set of capabilities that matter more for anyone doing actual knowledge work: computer and browser control, video editing assistance, long context handling, and the ability to generate working software from plain-language instructions. Early testers used it to edit real video projects in Final Cut Pro, comparison-shop on eBay, build slide decks, and run multi-step research tasks inside a browser, recording its own screen as it went.
TL;DR
- Computer use benchmarks for Astra reportedly beat prior models by a wide margin, and testers demonstrated it controlling a real browser to complete multi-step tasks like eBay research and drawing out workflows in Excalidraw.
- Long context handling appears meaningfully improved, with the model retaining information across much larger inputs instead of losing track partway through, according to benchmark results referenced in the model’s release材料.
- Video editing work was demonstrated directly, with Astra given Final Cut Pro access to import files, color grade footage, sync clips, and select the best audio track from several options, going beyond the original instructions in at least one case.
- A six-minute educational video was generated from a single prompt (“create a 5-minute educational video about T-cells”) and testers noted it lacked the obvious rendering artifacts that usually mark AI-made video as low quality.
- Astra completed Pokémon Fire Red in 18 hours on its highest settings, down from 96 hours for a prior GPT model and roughly 200 hours for the model before that, compared to a typical human completion time of 25 to 100 hours depending on play style.
- A Box AI evaluation of Astra on “complex work” tasks showed a 3% improvement on the full data set, with much larger jumps in specific industry subsets: technology moved from 62 to 77, legal from 64 to 72, and energy from 77 to 86.
- The model has a recognizable creative “smell,” repeatedly defaulting to forest green color palettes and flat design patterns across unrelated projects unless explicitly steered otherwise.
How does Astra handle browser and computer control?
Browser and computer use is where Astra’s improvements translate most directly into office work. One tester described giving Astra a task, telling it to complete the task inside the browser, and having it record its own screen (writing custom recording software rather than relying on a tool like QuickTime) while a timer ran in the corner. In one example, Astra searched eBay for three high-value Pokémon card listings, compared them, and finished the full research and comparison task in under two minutes.
This matters because it points at a different kind of usefulness than the 3D world-building demos. Booking appointments, comparing prices across listings, filling out forms, and pulling together research from multiple browser tabs are exactly the repetitive digital tasks that knowledge workers want off their plate. The demos shown so far are simple compared to a full workday of browser tasks, but they show the model completing multi-step sequences without falling apart partway through, which has historically been the failure point for “agentic” browser tools.
Is Astra good at video editing?
Testers gave Astra a limited, well-defined video editing job inside Final Cut Pro: import certain files, apply color grading, and sync multiple clips so a screen recording lined up with a talking-head video. This is a standard workflow for anyone producing tutorial or educational content. Astra completed the import and organization step by creating a separate folder structure on its own, without being asked, which surprised the testers running the demo. It handled the syncing correctly and did a workable job on color grading. It also went a step further, identifying which of several audio tracks was highest quality and deleting the rest.
None of this amounts to full creative editing. Astra wasn’t asked to make judgment calls about pacing, cuts, or story structure, and testers were clear that this was a scoped, mechanical task rather than a real editing test. But it’s a meaningfully different result than what prior models produced on video work, where reliability was consistently too low to trust with real footage.
Separately, one demonstration had Astra generate a full educational video from a single text prompt describing the topic (T-cells) and desired length. Testers noted the result looked cleaner than typical AI-generated explainer video, which usually carries visible artifacts or inconsistencies that flag it as machine-made. That threshold, output that doesn’t immediately read as “AI slop,” is what determines whether people will actually start using a tool for real output instead of just testing it.
What do the benchmark numbers say about knowledge work?
Seven tools to build an app. Or just Remy.
Editor, preview, AI agents, deploy — all in one tab. Nothing to install.
Alongside the demo videos, Box AI released its own evaluation of Astra on complex work tasks. Across the full evaluation set, Astra showed a 3% improvement over the comparison baseline. The more interesting numbers sit inside specific industry categories: technology tasks improved from a 62 to a 77 score, legal tasks moved from 64 to 72, and energy sector tasks jumped from 77 to 86. Media and entertainment also saw a large improvement, while consumer products saw a smaller, single-digit bump. These figures come from an evaluation built around document-heavy enterprise work, the kind of task where a model needs to extract, analyze, and reason over large sets of company files rather than generate creative output.
This lines up with the long context claims tied to Astra’s release. Long documents and large codebases have historically been a weak point for LLMs: models would claim support for very large context windows but effectively lose track of earlier information well before hitting the stated limit. Astra’s benchmark results suggest it retains relevant information across much longer inputs than earlier models, which is directly relevant to any workflow involving contracts, research papers, technical documentation, or large codebases.
Where does Astra still fall short?
The clearest limitation testers pointed to isn’t capability, it’s originality. Across five or six separate generated projects (websites, apps, visual demos) Astra defaulted to the same visual language repeatedly: forest green color schemes, flat design elements, and similar layout choices, even when given no design instructions at all. One tester called this an “AI smell” in both writing and design choices. The upside is that the model responded well to being steered toward a different look when explicitly prompted, suggesting the default output is a starting point rather than a hard ceiling.
The video editing and browser control demos were also narrow in scope. Editing a video by following an established process (import, sync, color grade) is different from making creative decisions about what footage to keep or cut. Browser tasks like comparing eBay listings are simpler than the multi-hour, multi-application workflows that make up a real workday. The underlying reliability looks improved, but the public demos so far test discrete tasks rather than sustained, open-ended work.
Frequently Asked Questions
What is GPT-6 Astra used for in real work settings?
Early testers used it for browser automation tasks like comparison research, video editing support inside Final Cut Pro (importing, syncing, color grading, and audio selection), generating slide decks, and producing full educational videos from a single text prompt.
How good is Astra at browser and computer control?
Testers reported it completing multi-step browser tasks, such as researching and comparing product listings, in under two minutes, and it reportedly scored well on computer-use benchmarks tied to its release. It also handled tasks like recording its own screen while working by writing custom software to do so.
Can Astra actually edit video, or just generate it?
Both were demonstrated. Testers gave it a defined video editing job in Final Cut Pro (import, sync, color grade) which it completed and extended on its own, and separately had it generate a full six-minute educational video from a single prompt.
Does Astra have any consistent weaknesses?
Yes. Testers noted it repeatedly defaults to similar visual styles (forest green palettes, flat design) across unrelated projects unless specifically directed otherwise, and the video and browser demos tested narrow, well-scoped tasks rather than long, open-ended work.
What do the Box AI benchmark numbers show?
Box AI’s complex work evaluation showed a 3% overall improvement for Astra, with larger gains in specific sectors: technology rose from 62 to 77, legal from 64 to 72, and energy from 77 to 86.
