OpenAI's GPT-6 Astra played Valve's Portal from start to finish with no human at the controls, spending $571.18 in tokens across 24 hours.
Key Points:
- GPT-6 Astra cleared every Portal test chamber on its own, with no person touching the controls at any point.
- The playthrough consumed 3,336 tool calls and $571.18 in API costs, covered by a $200 monthly Codex Pro plan.
- A local server fed the model screenshots and coordinates while the game stayed frozen during each reasoning step.
Portal Playthrough Logged 3,336 Tool Calls
An independent experimenter who posts under the handle cozyblaze wired the 2007 puzzle game into the model, then stepped back and let it work alone. Reporting on the attempt logged 3,336 tool calls and $571.18 in API charges at list rates. Astra cleared every test chamber, the escape through the facility's back rooms and the closing confrontation with the GLaDOS computer.
A local Model Context Protocol server sat between the model and a modified build of SourcePauseTool, a utility that Source engine speedrunners have relied on for years.
Portal froze while Astra read screenshots and player coordinates, then resumed for a set number of ticks so the agent could run its planned inputs.
The full session ran for 23 hours and 43 minutes, a stretch that included service interruptions and long periods of idle waiting. Actual gameplay filled roughly two hours once the thinking pauses were cut from the video. cozyblaze began the run at maximum reasoning effort, then resumed the session in Fast mode after repeatedly hitting capacity errors.
Also Read: Banks Sell Crypto While Rejecting Balance Sheet Risk In Brazil
Brockman Frames Astra Around Computer Use
The company released Astra on Sept. 3 and built the launch around computer use, meaning the model operates ordinary software much the way a person would. President Greg Brockman told reporters it moves through spreadsheets, forms and web pages at superhuman speed.
cozyblaze cautioned against reading too much into the run, calling it something short of a proper benchmark with plenty of problems left unsolved. Watching a general-purpose agent find its own way through an unfamiliar three-dimensional space still counted as real progress, the experimenter said. OpenAI listed a single agent solving a wide variety of games among its technical goals in 2016.
Astra Pricing Set The $571 Bill
Astra is priced at $10 per million input tokens and $50 per million output tokens, roughly 2.5 times the rate of the model it replaced. The $571.18 figure reflects OpenAI's list rates, though a $200 monthly subscription absorbed the bill.
Video games have proved a rough proving ground for large language models. An OpenAI model lost a beginner chess match to an Atari 2600 engine last year, and Anthropic's Claude spent weeks stuck inside a Pokemon game on a Twitch stream. The Portal run turns that record around, though the size of the bill shows how far the economics of agent work still have to travel.
Read Next: Robinhood Chain Flips Solana For The First Time With $1.45B In Daily DEX Volume
