PRODUCT UPDATE / Coding agents · 3D browser prototypes
OpenAI launches GPT-6.2 Astra and Sol with playable 3D demos
Both models are available in Codex and the API. OpenAI’s Astra demos show how developers can turn a brief into a walkable space or a working marble run.

Compare the two models on one scene or mechanic with the same playtest criteria.
OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on November 17, 2026, in Codex, ChatGPT Work and the OpenAI API. OpenAI positions Astra for harder coding and design tasks, and Sol for everyday coding and agent workflows. The release comes with playable browser examples, including a pinball table and a futuristic racer. OpenAI announcement · API changelog
Build a scene you can walk through
In Paris Atlas, OpenAI’s Thomas Ricouard used Astra with Codex to bring Blender landmarks into Three.js. That first build let him walk and drive through the city. Later passes refined details such as shopfronts, where the team checked the result across the city after fixing doors and windows that crossed storefront boundaries.
Ricouard’s Notre-Dame demo takes the same idea into one interior. For the reflection between its benches, the project uses a second camera mirrored below the floor. A walkable build lets a level designer judge that effect from the player’s height and see how it changes as the player moves.
Try a mechanic with moving parts
Escapement, an Astra demo by OpenAI’s Eric Provencher, follows marbles through a chain of mechanisms. The project record starts with a small wooden track and shows successive additions, including a tipping platform and sound effects triggered on contact. A game team could use that approach to test whether one trap or puzzle mechanism resets cleanly before building a whole level around it.
Choose a model for the first playable build
In the API, the model IDs are gpt-6.2-astra and gpt-6.2-sol. At launch, standard input and output cost $8 and $40 per million tokens for Astra, and $2 and $10 for Sol. Cached input costs $0.40 and $0.10 per million respectively. Access in Codex and ChatGPT Work depends on the account and, where applicable, workspace settings. Launch pricing · ChatGPT Learn
The three projects shown here use Astra. For a comparison in Codex, give both models the same short brief and run the builds in the same browser. Check whether the camera behaves along the route, the mechanic works after a reset and the result survives a change to the brief. That gives you a concrete result to review alongside the time and tokens each build took.
Sources (7)
Based on OpenAI’s release and public demos. No sponsorship or affiliate links.




