GPT-6 Astra Real-World Use Cases: What People Built First

Two weeks in, the most useful signal on GPT-6 Astra isn't the benchmarks — it's the flood of documented projects people have actually shipped with it. This roundup collects the verifiable ones, with sources, and the costs and failure modes those same users report.

Sourcing note: the centerpiece here is a widely shared community catalog of 366 tweeted Astra projects (compiled September 13), plus full Hacker News threads, and published hands-on reviews. The catalog's own compiler warns these are "reports in retained tweets, not independently verified successes or production deployments" — we've kept that caveat attached and favored writeups with reproducible detail.

The 366-project catalog: what people are building

The category counts are themselves the story of what Astra changed: 73 architecture and interior-design builds, 71 playable games, 51 interactive 3D models, 40 video-editing and motion projects, 30 CAD and physical-product designs, 30 websites, 26 illustrations, and 26 software-repair and workflow automations. The 3D-heavy skew is new — no prior model launch produced this distribution, and it matches early testers calling spatial modeling Astra's sleeper capability.

Standouts with public documentation: a Manhattan recreated street-by-street in Unreal Engine over a week using a manager/implementer agent workflow; Zillow listing photos turned into a 3D house model plus promo video in one pass; a 38-page blueprint PDF converted to a to-scale walkthrough "in about ten minutes"; a telescope star-tracker PCB designed, optimized, and actually ordered; a five-story Revit building with a full PDF sheet set generated from a single image; a custom LEGO Technic truck with 137 assembly steps; and an open-world RPG, "Morta's Blade," that reached the App Store. On the computer-use side, agents completed all 48 levels of the "I'm Not a Robot" CAPTCHA game and finished Portal.

Computer use in practice: DaVinci Resolve, Xero, Figma

The most instructive writeups are the mundane ones. The AI Automators published six worked computer-use examples with costs and timings: Astra edited B-roll into a DaVinci Resolve timeline (about 20 minutes elapsed for what it estimated as "31 minutes of work"), reconciled seven bank transactions in Xero in about ten minutes, built a 3D logo by writing the Blender file directly instead of clicking through menus, and beat a browser chess bot by spinning up a local Python engine and clicking its moves in. Their cost reality-check: those six examples consumed roughly 40 percent of a $100/month plan's usage. ChatPRD's Claire Vo reported the first agent success she's seen inside Figma — YouTube thumbnail production end-to-end — and said Astra "one-shotted a bunch" of tasks that had failed for months on earlier models.

Long-horizon coding: week-long Codex runs

The heaviest documented workloads are multi-day autonomous coding runs. One HN developer runs seven Codex 20x accounts across three machines on game decompilation with 50 sub-agents — a single 80KB function took "about a week of constant sol-ultra attention." Another had Codex convert OpenStreetMap data into his VW's proprietary offline-map format in "around a week... with some light steering" — work he estimated would have taken him months. Vincent Schmalbach ran Astra as his Codex driver while on vacation, giving it about an hour of instruction per day and letting it run autonomously — and burned through two $200 Codex subscriptions in one week. His core criticism doubles as the capability's description: "It still does not know when to stop."

Professional workflows beyond code

Outside engineering, reviewer Matthew Berman — whose verdict was "This is the best model I've ever used. Period." — highlighted slide decks, research, and browser control "significantly faster" than GPT-5.6, plus a simulation project that ran five days straight designing each asset individually. OpenAI's own launch demos leaned professional: filling a Form 1040, job searching, and formatting legal documents in LibreOffice. Legal-tech company Legora's 41-documents-in-minutes review workflow is covered in our earlier report.

The costs and caveats users report

Three consistent warnings from the same people shipping these projects. First, usage burn: multiple developers measured Astra consuming subscription limits at about 2.5x the rate of GPT-5.6 Sol — the flip side of its per-task token efficiency is that people hand it much bigger tasks. Second, stopping behavior is broken in both directions: some report tasks spiraling into 12-24 hour runs with excessive verification, while others had it stop four times mid-task asking for permission already given — OpenAI's own docs acknowledge both and recommend defining completion criteria up front (the exact prompt language is in our best practices guide and the free PDF above). Third, safety stops are real: Astra is the first model at OpenAI's Critical cybersecurity threshold, its misalignment monitoring can kill a conversation irrecoverably mid-run, and production guides recommend treating those stops as terminal states excluded from retry logic. Plan for all three before you point it at anything that matters.