Your game freezes mid-frame. An AI model analyzes a screenshot, reads the player’s position, decides where to fire a portal, and unpauses for a fraction of a second. Then pauses again. Thinks again. Decides again. Repeat that loop 3,336 times across 23 hours and 43 minutes, and you’ve got what independent developer cozyblaze built — OpenAI’s GPT-6 Astra autonomously completing Valve’s Portal, end credits included, for a headline API cost of $571.18. Milestone or expensive flex? Honestly, it’s both.
How Astra Actually Played the Game
The pause-think-act loop ran 3,336 times before Portal’s end credits rolled.
Cozyblaze connected Astra to Portal using Model Context Protocol — MCP, a framework for wiring AI models to external software — alongside a modified SourcePauseTool. The game pauses, Astra receives a screenshot and positional data, makes a decision, and the game unpauses to execute those inputs. The condensed video of the run clocks roughly two hours. The actual session ran nearly 24.
The operational numbers tell the real story:
- 23 hours, 43 minutes total wall-clock runtime
- 3,336 tool calls made by Astra throughout
- $571.18 in API tokens at list pricing ($10 per million input tokens, $50 per million output tokens, per specialized tech reporting)
- Cozyblaze’s actual out-of-pocket cost differed — they hold a $200/month Codex Pro subscription
- The condensed gameplay video trimmed all AI “thinking” pauses down to about two hours
Tokens are the units of text and image data the model processes each cycle. With 3,336 decision loops — each requiring a fresh screenshot analysis and positional read — the bill adds up fast.
What This Actually Proves (And What It Doesn’t)
Playing a well-documented game with borrowed knowledge isn’t the same as solving an unfamiliar one from scratch.
Portal has been exhaustively documented online for nearly two decades — walkthroughs, wikis, frame-perfect speedrun guides. Astra leaned heavily on all of it. Cozyblaze characterized the run as a proof of concept for multimodal task orchestration, not a benchmark of general game intelligence. What you’re watching isn’t an AI reasoning through an unfamiliar puzzle; it’s a generalist model cross-referencing internet knowledge while controlling software through vision and tool calls.
That distinction matters more than the stunt itself. DeepMind’s AlphaGo and StarCraft RL agents were specialists — trained exhaustively for one domain. Astra is a general-purpose model that improvised its way through a 2007 cult classic using borrowed knowledge and screenshots. The shift from specialist to generalist is real, even if the cake — as Portal fans know — is a lie. The Verge noted, somewhat dryly, that the run “probably only required a small swimming pool’s worth of water” in data-center cooling costs, a qualitative reminder that the $571 API bill is only part of the tab.
What Comes Next
Right now it reads as a party trick, but the architecture it’s rehearsing is considerably less trivial.
As token prices drop, setups like cozyblaze’s point toward genuinely useful applications — automated QA testing for complex software, accessibility tools, AI coaching systems. Spending $571 to autonomously complete a puzzle game remains firmly in proof-of-concept territory. But the underlying architecture — a general model, vision input, tool calls, long-horizon task completion — is the same scaffold future AI agents will build on. Cozyblaze just assembled the early prototype. Rough edges and all.
