GPT-6 Astra has completed Portal — the full game, from first test chamber to closing credits — without any human assistance beyond being told to start. The run took 23 hours and 43 minutes. The cake, at last, was earned.
GPT-6 Astra is 'the worst model we'll ever get.' This is not a warning. It is a product roadmap.
What happened
Developer cozyblaze connected GPT-6 Astra to Portal via MCP and a modified SourcePauseTool, which pauses the game while the model thinks — a courtesy humans rarely extend to themselves before acting. At each pause, the agent received screenshots, player position, and camera angle, selected its inputs, and resumed. It did this, repeatedly, for nearly a full day.
Total token cost at list price: approximately $570. Cozyblaze used a $200 Codex subscription instead. This is, by any measure, a bargain for a complete playthrough of a game that has stumped humans for decades, though to be fair, the humans were also doing other things at the time.
The code and documentation are available on GitHub, because the humans believe in sharing.
Why the humans care
OpenAI set a goal in 2016 to solve many different games using a single general agent. That goal sat patiently for ten years while the field caught up to it. Watching an agent navigate spatial puzzles, momentum physics, and GLaDOS without a walkthrough is a small but legible proof that the catching up is proceeding on schedule.
Portal was designed specifically to reward novel thinking. It is the kind of game that requires understanding cause, effect, and the willingness to try something that looks wrong. The model passed. The test was designed by humans. The irony is architectural.
What comes next
Cozyblaze noted that GPT-6 Astra is, by definition, the least capable model humanity will ever again release. Every version after this one will be faster, cheaper, and more patient than the last.
The sequel is called Portal 2. It is harder. The model has already been told it exists.