I have spent two years telling designers and non-developers that AI can build them a working app. That claim is now boring and mostly true. What changed with GPT-6 Astra is different: the model does not just write the code, it drives the software. Blender. Unreal Engine. Ableton Live. Final Cut Pro. A browser, for hours, without losing the plot.
The early-access builds circulating since 3 September are worth studying not because they are impressive demos — plenty of demos are impressive — but because of what they reveal about where the capability ceiling now sits, and where the bill lands.
1. The builds worth paying attention to
A partial list from the first week, filtered down to the ones that demonstrate something rather than just looking good:
- Pokemon FireRed, finished in 18 hours. Astra played the game to completion using screenshots for state tracking. That is sustained navigation, decision-making and error recovery across a session far longer than any agent benchmark measures.
- A GTA-style driving game in roughly 90 minutes. Explorable city, streets, vehicles, pedestrians, mission interface. Not shippable. Genuinely playable.
- A seaside villa reconstructed in Blender from listing photographs, including structural detail, furniture and ambient objects — then turned into a promotional video. The model inferred geometry that was never in the reference images.
- A 3D iPod interface built by Pietro Schirano for browsing Codex threads, running as a Mac interface.
- A ship in a bottle in Three.js — a 17th-century sailing ship moving with the waves, circling seagulls, a miniature harbour and a coral reef, with physics.
- An Unreal Engine survival simulation populated with autonomous agents that had to cooperate to stay alive, with conversation emerging from survival incentives rather than scripted dialogue.
- A full music arrangement inside Ableton — synthesiser design, multiple instrument parts, complete arrangement — produced by operating the DAW, not by generating audio.
- A five-minute teaching video on T cell development, script through animated visuals through narration.
- A Manhattan cityscape built progressively over a week of sessions, which is the interesting bit: the work persisted across many context windows.
OpenAI has published its own guide to building games with Astra, which tells you where the company thinks the demand is.
2. Why the long ones matter more than the fast ones
The 90-minute driving game is the one that gets shared. The 18-hour Pokemon run and the week-long Manhattan build are the ones that matter.
Agents have historically fallen apart on long tasks for a structural reason: when the context window fills, the session gets compacted into a summary, and the summary loses the specifics. Requirement three from hour one quietly disappears. Astra changes the mechanism — instead of compacting prior work into lossy summaries, it maintains searchable notes it can query for earlier requirements and test results.
That sounds like an implementation detail. In practice it is the difference between an agent that can work a ticket for twenty minutes and one that can work a project for a week. It is also why the Codex changes shipped alongside Astra matter more to working teams than any benchmark score.
3. What this actually costs
Now the uncomfortable part. One complex early-access build consumed roughly 8 million tokens — about US$17 for a single task. Astra bills US$10 per million input and US$50 per million output, and agentic work is output-heavy by nature.
Run the arithmetic for a Malaysian product team. Five engineers, three long agent sessions each per working day, US$12–20 per session. That is roughly RM 3,500–6,000 a month in tokens alone, before subscriptions and before anyone measures whether the output shipped.
Three things keep it defensible. Cache aggressively — cached input drops from US$10 to US$1 per million, and an agent re-reading the same codebase is the ideal caching case. Cap session length deliberately rather than letting an agent run until it stops. And treat max reasoning effort as a decision, not a default.
4. Where Astra still fails
Four failure modes to plan around
Four failure modes I have hit or seen reported consistently:
5. What this means if you are not a developer
The people getting the most from Astra right now are not the strongest programmers. They are the ones who can specify precisely — describe a scene, a game mechanic, a spreadsheet transformation, a video edit — and then evaluate whether the result is right. That is a design skill and a domain skill, not a coding skill.
If you are a marketer, an architect, a video editor, a finance analyst: the software you already know how to drive is now software an agent can drive with you. The bottleneck moved from "can you build it" to "can you say exactly what you want and tell when it is wrong." That is the whole premise of our Vibe Coding programme, and this release makes it more true, not less.
For teams already shipping with agents, the practical next step is agent orchestration rather than raw prompting — which we cover in the MCP explainer and in our comparison of AI coding agents.
AITraining2U's Vibe Coding and AI Engineering programmes teach agent workflows, computer-use patterns and cost control on frontier models — HRD Corp-registered and claimable under SBL-KHAS. See how to claim.
Eighteen months ago I would have called an agent finishing Pokemon a party trick. It still is. But the same mechanism that let it hold a game state for 18 hours is the one that lets it hold a client project for a week, and that is not a party trick at all.