Earlier this month, an enthusiast going by CozyBlaze hooked OpenAI’s new flagship model, GPT-6 Astra, up to Valve’s puzzle game Portal and let it play from start to finish with zero human input. No walkthroughs typed in mid-session, no hints, no manual steering — just screenshots, a paused game clock to “think” in, and a controller it operated itself. It finished the whole game.
The headline number doing the rounds is that this cost somewhere north of $57,000 in tokens. That figure is wrong — it’s a misread of the article URL, which drops the decimal point. The actual cost was $571.18. Still not nothing, but two orders of magnitude away from the number people are quoting at each other in comment sections.
Once you get the number right, though, the story underneath it is genuinely worth ten minutes of an IT operator’s time — not because a chatbot beat a fifteen-year-old game, but because of what the mechanics reveal about where agentic AI actually is right now, and what it costs to run.
What Actually Happened
Astra controlled Portal through the Model Context Protocol (MCP), paired with a modified version of a tool called SourcePauseTool that let it pause the game engine while it worked things out. The loop was simple in concept: the game freezes, Astra gets a screenshot and the player’s position, it reasons about the puzzle in front of it, plans a sequence of moves, then unpauses and executes them. Repeat, for roughly 24 hours of wall-clock time (edited down to about two hours for the published video).
That’s 3,336 individual tool calls — each one a screenshot in, a decision out — to navigate 3D space, understand spatial puzzles involving portals and momentum, and chain together a full playthrough without a human course-correcting it.
This isn’t a new party trick. OpenAI has talked since 2016 about building a general-purpose agent that can play a wide range of games, and this is a fairly direct descendant of that ambition — an agent given a visual interface and a goal, left to work out the “how” on its own.
The Bit That Deserves an Asterisk
Before anyone gets too impressed, it’s worth being straight about the caveat that a lot of the coverage buried: Portal has been extensively documented online for over a decade. Strategy guides, speedrun routes, full puzzle walkthroughs — all of it is almost certainly somewhere in Astra’s training data. So this wasn’t an AI reasoning through entirely novel spatial puzzles from first principles the way a human player encountering Portal for the first time would. It was, at least in part, pattern-matching against things it had already “read” about the game.
That doesn’t make the demo meaningless — controlling a game engine, sequencing dozens of precise physical actions, and recovering from its own mistakes across 24 hours without falling over is still a real display of long-horizon tool use. But it’s a good reminder to read agentic AI demos the way you’d read a vendor benchmark: impressive, real, and not quite what the headline implies.
The Number That Actually Matters for Your Business
Here’s the detail I think is more useful than the game itself: CozyBlaze didn’t pay $571 out of pocket for this. It was covered under a $200-a-month Codex Pro subscription — a flat-rate plan that happened to have enough headroom to absorb over 3,000 tool calls and a day of continuous agentic activity.
That’s the part worth sitting with if you’re evaluating agentic AI tools for your own operation. Long-horizon agents — the kind that take a goal, break it into steps, and keep working autonomously until they’re done — can rack up token usage fast, and the relationship between “one task” and “one bill” isn’t always obvious upfront. A single request that spawns thousands of tool calls behind the scenes looks nothing like a normal chatbot conversation on your invoice.
What This Means If You’re Running a Business
None of this is a reason to avoid agentic AI. It’s a reason to go in with your eyes open about how it’s priced and governed.
Ask what “autonomous” actually costs before you deploy it. If a tool or workflow is going to plan its own steps and call other tools repeatedly to get there, get a sense of typical token consumption per task before you roll it out at scale — not after the first invoice.
Favour subscription or capped-usage plans for exploratory agentic work. Flat-rate plans like the one in this story exist precisely because usage-based agentic billing is hard to predict. If you’re trialling this kind of capability internally, a capped plan limits your downside while your team learns what “normal” usage looks like.
Treat capability demos with a healthy amount of scepticism. When you see an AI vendor showcase something impressive, ask what’s actually novel versus what’s likely baked into training data. It changes how much weight you should put on the demo when deciding whether the same capability will hold up on a task specific to your business that nobody’s ever written a guide for.
Long-horizon autonomy still needs a human checkpoint. A 24-hour unsupervised run makes for a great video. In a business context, that’s also 24 hours where nobody’s watching what an agent is doing with real systems, real data, or real customer information. Build in review points, not just a start button.
The Bottom Line
An AI playing a video game unassisted is a fun story. The actual lesson for anyone running a business is about cost transparency and governance around agentic AI — tools that don’t just answer a question but take a goal and run with it. That’s a genuinely useful capability, and it’s also one that needs a proper cost model and some guardrails before it touches anything that matters to your business.
If you’re trying to work out whether agentic AI tools are worth trialling in your business — and what it would actually cost and require to do it safely — get in touch or have a look at what we offer.