One prompt, eight games. GPT-6.2 Sol and GPT-6.2 Astra at four effort levels.
OpenAI's GPT-6.2 Sol and GPT-6.2 Astra, released on 17.11.2026, each built a first-person voxel sandbox game in Codex CLI from the same prompt and no picture: a generated block world to walk through, mine and build in. One run each on low, medium, high and xhigh effort, with no human edits.
This is not a CmdBench benchmark run: no checks, scores or verdicts.
The eight games
One run per model and effort level. Each game is served exactly as the agent wrote it, with no human edits, and opens full screen in this tab.
GPT-6.2 Sol, low effort
Fieldcraft: 80 × 80 blocks, drawn with Three.js in one script, with five blocks to build with.
Click the game to look around; each game shows its own controls. Made for a mouse and keyboard. Opens full screen in this tab.
The same scripted minute of play, in four frames
0:02Before the first click
0:19Mining
0:33Building
0:59The sky, later
In the same scripted minute of play, all eight games took the mouse and keys: each one locks the mouse on a click and looks around with it, mines blocks, picks a block by number key and by mouse wheel, places blocks, and changes its sky as the minute passes. Brinewild, GPT-6.2 Astra's game on medium effort, starts at dusk, so most of its minute is at night and its frames are dark.
Codex CLI can't be limited to file tools, so every run could also run commands in a sandbox (its own folder, no network, no browser). Each run's log shows it ran none; it only wrote files.
Costs are estimates at API list prices. Codex CLI reports no cost, so each is computed from the run's own token counts at OpenAI's list prices of 17.11.2026.
One run per model and effort level: exploratory, not a ranking.
GPT-6.2 Sol (gpt-6.2-sol), GPT-6.2 Astra (gpt-6.2-astra)
Agent
Codex CLI 0.169.0
Effort levels
low, medium, high and xhigh, one run per model at each
Tools
Codex CLI can't be limited to file tools, so each run could also run commands in a sandbox: its own folder only, no network, no browser. Each run's log shows it ran none; it only wrote files. No model saw its game running.
Time limit
60 minutes per run
Three.js
0.187.0, in the build folder: the only library the prompt allowed
Human edits
None. Each game is served exactly as the agent wrote it.