OpenAI's newest model finished Portal alone in about 24 hours.
A developer let GPT-6 Astra, OpenAI's newest model, play Valve's puzzle game Portal with no human help. It pressed keys like a person and froze the game while it reasoned, taking about 3,300 tool calls. The run cost $571 in tokens, and the developer refuses to call it a benchmark.
Why it mattersPlaying a game by sight and keypress is the same skill as running your computer. Portal is also one of the most documented games online, so Astra had likely studied it.
Read in the app — free, in 9 languages
Related stories
Most Americans told pollsters that AI is bad for society.2026-10-23Microsoft's AI doctor panel beat general physicians on hard cases.
2026-10-23Anthropic gave away control of its Agent Skills standard.
2026-10-23Outside testers found ways for OpenAI's agents to hide their actions.
2026-10-23Walmart opened its checkout to AI agents that shop for you.
2026-10-22