Astra Is Here. Naturally, Someone Made a Pelican Ride a Bike

Astra Is Here. Naturally, Someone Made a Pelican Ride a Bike

GPT-6 Astra arrived on 3 September.[2] Within a day, someone had used it to make a low-poly pelican ride a bicycle in Blender.[6]

This is how you know a new technology has entered the useful stage. The launch presentation may be full of benchmarks, professional workflows and grave statements about intelligence. The users immediately ask it to build a cycling seabird.

OpenAI describes Astra as its strongest model for computer use, browsing, software engineering, science and professional work. It is rolling out to ChatGPT Plus, Pro, Business and Enterprise, alongside API and cloud access.[1] The interesting change is visible in the early experiments: people are giving it an outcome, access to familiar software and enough time to figure out the messy middle.

The results range from genuinely practical to gloriously unnecessary. Both categories are worth exploring. The launch film also drew comparisons with WALL-E and Her, as viewers debated whether handing over every computer task looks liberating or a little too comfortable.[9]

The official demo went from a yellow circle to a rocket ship

OpenAI’s launch film begins with someone drawing a small yellow circle. Astra turns it into the window of a rocket, adds detail, opens Blender and creates a 3D model. The same demo asks it to build an asteroid-dodging game, prepare a colourful retail presentation, create an eBay listing for a dented table, find a tennis court and prepare a 3D-printable file.[2]

That sequence explains Astra better than a benchmark chart. The model can carry an idea through several forms. A sketch becomes a design, the design becomes a 3D object, and the object becomes something printable or playable.

It also raises a useful question for the rest of us: what half-finished idea is sitting on your phone because turning it into something real requires three unfamiliar applications?

1. Build the silly game you would have played as a child

Games dominated the first wave of Astra experiments.

Pietro Schirano shared an underwater exploration game made from one goal prompt, with the 3D elements and sounds generated from scratch. He also made a PS1-style spinning-top game and used Astra with Ableton to create a music track from synthesised instruments.[5]

Matthew Berman showed a one-prompt obstacle game with sound effects, a walkable city made from ASCII characters, a tiny planet, a liminal-space horror game, a multiplayer puzzle, and a five-day attempt at a working city simulator. The city project was still unfinished when he decided to ship it, which is both a limitation and possibly the most authentic game-development result imaginable.[3]

Playco took a more structured route. It used Astra inside an AI development environment connected to Unity and Godot. From one basic grey-box game, the team produced three themed prototypes and reported 50% fewer manual fixes than with the previous model.[7]

Try this:

Build a small browser game inspired by a playground game I loved as a child. Start by asking me five questions about the rules, mood and art style. Make one level, include sound controls, test the win and lose states, and give me a playable preview. Keep the scope small enough to finish today.

The last sentence matters. Astra can work for a long time, but “make my dream game” is still an excellent way to get a large, expensive pile of features.

2. Turn your living room into a 3D scene

Dominik Kundel gave Astra ten quick photographs of a home bar and asked it to rebuild the room in Blender. He supplied no measurements. Astra used recognisable objects in the photographs to estimate scale, then modelled the room, shelves, bottles, furniture, a chair and even stray wires.[4]

The appeal goes beyond showing off a 3D render. You could use the same approach to plan a room, test a different layout, preserve a favourite place as a digital memory or create a background for a family game.

Try this:

Use these photographs to create a rough 3D model of this room. Do not invent exact measurements. Explain which objects you are using as scale references, label uncertain dimensions, and build a first pass with simple materials. Let me review the layout before you add detail.

That prompt asks for uncertainty to remain visible. A convincing render can still contain a sofa that would not fit through the door.

3. Make the LEGO set that does not exist

Kundel also gave Astra access to BrickLink Studio and asked for the Golden Gate Bridge. About ten minutes later, he had a first version with small cars on the road. Astra worked out which parts existed in the application’s bundled library, assembled the model with Python, opened it in Studio, inspected the result and continued refining it.[4]

He later asked Astra to build a site where people could rotate the models in 3D and download the build files. It added an authenticated upload area for the owner.[4]

This is a lovely example because it crosses several jobs: research the available pieces, design the object, inspect it in specialist software, then package it for other people.

Try this:

Design a buildable brick model of a local landmark using parts available in BrickLink Studio. Keep it under 500 pieces. Create a first version, inspect it from every angle, check for unsupported sections, and produce a parts list. Ask before substituting any rare or expensive piece.

A family could do the same with a dream treehouse, a pet, a football ground or a grandparent’s old shop.

4. Walk through a painting—or through history

One early tester built a playable London that moved through Roman, Saxon, medieval and later eras. The same tester created environments based on famous paintings, including a world where the viewer could walk between Van Gogh scenes. He repeatedly warned that historical and scientific accuracy needed checking, and pointed out visible mistakes such as floating architecture and incorrect structures.[10]

Those caveats make the idea more useful, not less. Astra can create an explorable interpretation quickly. It cannot turn visual confidence into historical truth.

Try this:

Create a short interactive walk through three paintings by the same artist. Preserve the palette and major composition of each work without claiming the unseen areas are historically accurate. Add a source panel for every factual note. Clearly label generated interpretation, and let the visitor move between scenes through visual transitions inspired by the artist.

For education, the finished experience should be reviewed by someone who knows the subject. For fun, it can be a brilliant way to explore how a painting might feel beyond its frame.

5. Build an absurd simulator

One of Berman’s strangest projects placed five million pieces of hay in a digital haystack with one needle. The player can inspect them one by one.[3]

This may be the purest use of frontier intelligence yet discovered.

The same creative instinct can produce a virtual ant farm, a duck-run café, a village where every resident is convinced they are the detective, or a weather system controlled by the mood of a houseplant.

Try this:

Make a tiny simulation called “The Village That Misplaced Tuesday.” Give 20 residents simple routines and conflicting theories about where Tuesday went. Let me change one town rule at a time and watch the social consequences. Include a reset button and a readable event log.

Small simulations are good Astra tests because they combine rules, interface design, state and visual feedback. They also reveal quickly whether the model has built a system or painted a convincing façade.

6. Give your hobby its own piece of software

Astra’s strongest personal use may be software made for an audience of one.

You could create a garden planner that knows the odd shape of your beds. A football training tracker designed around your team. A searchable archive of family recipes. A tool that turns a child’s drawings into printable paper toys. A dashboard for a model railway. A holiday map where every family member pins one non-negotiable stop.

OpenAI says Astra can build and host websites, web apps and games through Sites in ChatGPT.[1] Kundel’s experience suggests the model can also think through interactions that were not specified in detail, although he warns that it may add too much when the user has a precise result in mind.[4]

Try this:

Build a private web app for my hobby. Interview me about the activity, the information I repeatedly lose and the task I find annoying. Propose three tiny tools, recommend the smallest useful one, then build only that. Use sample data first and ask before connecting accounts or publishing anything.

The goal is not to found a startup by teatime. It is to stop adapting a generic app to a very particular life.

7. Let Astra create across several applications

The recurring pattern in the best examples is movement between tools.

Schirano connected Astra to Ableton to build music.[5] Kundel used BrickLink Studio, Blender, video-editing tools and a browser.[4] OpenAI demonstrates Astra working across creative, office and engineering software, and says it can inspect its own websites with frontend QA.[1]

This suggests a fun weekend challenge: choose one object and make Astra express it in five forms.

Start with a child’s monster drawing. Turn it into a character sheet, a 3D model, a short theme tune, a tiny game and a printable card. Or take a family story and turn it into an illustrated map, an audio scene and a small interactive archive.

The result will not be equally good in every medium. That is the point. You are testing whether the model can carry an idea across tools without losing the thing that made it interesting.

Astra still needs an adult in the room

The launch has a serious edge. OpenAI classifies Astra at the Critical threshold for cybersecurity capability and is restricting its most advanced cyber functions. The company says additional monitoring may also pause legitimate long-running work.[8]

Everyday computer use carries smaller but immediate risks. A model that can list a table on eBay can enter the wrong condition, price or delivery details. A model that can book tennis can also choose the wrong time. A model building from personal files may see more than the task requires.

Use a simple rule: let Astra explore and prepare; keep approval for anything that spends money, publishes publicly, sends a message, changes an account or makes a consequential claim.

Early users also report familiar rough edges. Berman noticed repeated visual defaults across unrelated projects and some unfinished or overgrown builds.[3] Arena AI’s first-impressions test found impressive 3D scenes alongside floating bridges, incorrect structures and uneven interaction.[10] A polished surface remains evidence of good presentation, not proof that every detail works.

The best first experiment is one you abandoned

Kundel had previously spent hours trying to make an AI build LEGO models and concluded the models were not good at it. Astra produced a credible first bridge in about ten minutes.[4]

That is the most useful challenge for new users. Do not begin with “What can Astra do?” Open the folder containing the drawing, game, family archive, room plan, video or odd little tool you gave up on six months ago.

Give Astra the software you would use, a few good references, a clear finish line and permission to ask before important decisions. Then see whether the cycling pelican has company.

Sources

[1] https://openai.com/index/gpt-6-astra — GPT-6 Astra: A new generation of intelligence [2] https://www.youtube.com/watch?v=1QNsdr-Qx_I — Introducing GPT-6 Astra [3] https://www.youtube.com/watch?v=ZTgFyP0PZJo — I've had early access to Astra [4] https://x.com/dkundel/status/2095972046014673156 — Five things I learned from using Astra [5] https://x.com/skirano/status/2095595944762880070 — Pietro Schirano Astra tests [6] https://community.openai.com/t/astra-in-action-share-your-builds-breakthroughs-aha-moments/1394945 — Astra in Action community thread [7] https://openai.com/index/playco-game-prototyping-with-astra — Playco game prototyping with Astra [8] https://openai.com/index/path-to-astra — Path to Astra: critical capabilities and frontier safeguards [9] https://tech.yahoo.com/ai/chatgpt/articles/openais-ad-latest-model-looks-155340852.html — OpenAI's Astra ad draws WALL-E comparisons [10] https://www.youtube.com/watch?v=GQPi39sjNhU — GPT-6 Astra first impressions — Arena AI