Astra finished Portal in 23 hours and 43 minutes of wall clock
A hobbyist pointed OpenAI's new model at Valve's Portal on Friday evening and let it play to the credits. It got there in under a day, with nobody at the keyboard after the first instruction — and the number that should stick is not the completion, it's the bill.
GPT-6 Astra played the original Portal from the first test chamber to the end credits with no human help after the opening goal was set. The run, built and posted by a developer who goes by cozyblaze, started at 5:00pm San Francisco time on September 4 and finished 23 hours and 43 minutes later, and the code, setup docs and a sanitized 6,925-record session log are all public. Astra didn't get a game API — it got the same three tools a speedrunner would: read the API documentation, take a screenshot, and execute a plan. Between decisions the game freezes. A patched build of SourcePauseTool stops the simulation, hands the model a screenshot plus the player's position and camera angle, waits for it to pick camera turns and button presses with a duration for each, then advances up to 6,600 ticks — about 99 seconds of play — before pausing again.
The token accounting is the part that turns a stunt into a data point. Across the run the session logged 3,336 tool calls, 433 million input tokens and 1.6 million output tokens, of which 1.18 million were reasoning — that works out to roughly $571 at Astra's list price of $10 and $50 per million, though cozyblaze ran it on a $200 Codex subscription rather than paying the API bill. Call it what you like: a purchased copy of a nineteen-year-old game, or about a day of what the median OpenAI researcher now spends on inference by mid-afternoon.
Two caveats sit inside the developer's own notes, and they matter more than the headline. The human in the loop wasn't absent — cozyblaze resumed the session after capacity errors, switched the model to Fast mode, and issued one further instruction, to let the credits roll. And Portal is the most documented puzzle game on the internet, so this is not evidence that Astra can solve an unfamiliar game from scratch; even its builder calls it "not a proper benchmark per se." What it is evidence of is long-horizon bookkeeping: an agent holding a coherent goal across 3,000-plus decisions and 26,000 logged records without losing the thread. That is exactly the failure mode we wrote about this afternoon — OpenAI's own data says autonomy decays with the length of the task — and this is what it looks like when the thread holds.
The same capability is being sold as a compute story. Tae Kim, the Nvidia biographer, spent the weekend handing Astra a Mac laptop and a fresh Blender install — a program he had never used — and asked for a Space Shuttle. It researched the shuttle online, wrote its own scripts and drove the application itself, returning a rotatable 3D model in about ten minutes. His conclusion is an investor's one: computer use is the fourth exponential demand wave in four years, after chatbots, reasoning models and agentic coding, and each of the previous three sent orders for GPUs, high-bandwidth memory and CPUs vertical.
Brockman's launch-briefing line is the supply-side version of the same claim — computer use has "crossed the qualitative threshold," going from "slow, painful to watch" to "accurate, efficient, and reliable," so that Astra "can really do anything a human can do with a computer without an API." Kim adds the number that shows where the demand comes from: OpenAI's own post says the median company researcher went from negligible coding-agent use in January to more than $600 of tokens a day by mid-August, with the 90th percentile above $7,000.
The bull case has a testable shape, and it is the same test that Astra's benchmark numbers keep failing. Kim's thesis is demand-side and probably right — driving software that has no API is the largest reachable-surface expansion in agent history, and every one of those sessions re-reads the whole screen on every step, which is why 426 of Astra's 433 million input tokens in the Portal run were cached. But "demand wave" and "durable workload" are different claims. Portal cost $571 partly because the game was paused for 21 of the 24 hours while the model thought; a four-wave compute thesis that rests on people paying frontier rates for that ratio needs the cost per completed task to fall faster than the benchmarks currently suggest. We laid out that arithmetic when independent benchmarks put Astra five points behind Anthropic — the model is on the cost-per-task frontier for agentic coding and 75% more expensive per task than its predecessor everywhere else.
What to watch: whether anyone reruns the Portal setup on a game with no walkthroughs on the open web, and whether OpenAI reports hands-off completion rates — not success rates — for the multi-hour computer-use tasks it is now selling.
If an agent can finish a game nobody touched for 23 hours, what's the first piece of your own software you'd actually hand it — and what would the token bill have to be before you did? Tell us in the comments.
Sources: The Decoder — GPT-6 Astra beat Portal start to finish without human help in under 24 hours · portal-agent — how it works (GitHub) · portal-agent — session log and timeline (GitHub) · portal-agent — token usage summary (GitHub) · cozyblaze on X · The Verge — GPT-6 Astra beat Portal, and it only cost $571.18 · AI Weekly — GPT-6 Astra clears Portal autonomously for $571 in API costs · Key Context by Tae Kim — GPT-6 Astra and the Fourth Exponential · Techmeme
