AI Founder Weekly

Opus 5.5 built a 277k-gate computer in one 30-hour run

One prompt, no supervision, and a machine you can actually run in the browser.

Matt Shumer says he gave Claude Opus 5.5 a single prompt and it came back about 30 hours later with a working computer built from scratch in code. His count is 277,000 logic gates written in JavaScript, with a CPU, memory, an assembler, a tiny operating system on top, and games running on that OS. He posted a link to run it and says the whole thing was done autonomously in ultracode mode.

Matt Shumer
@mattshumer_
X
This isn't a mockup... it's a computer.
Sep 27, 2026 · View on X
Matt Shumer
@mattshumer_
X
treat it like you're proving you could have invented computing yourself. go all out.
Sep 27, 2026 · View on X

The prompt is the interesting part, and Shumer published it in full. He asked for the game to actually run on the gates, "not in javascript pretending", and for the whole machine to be visualised in 3D so he could play the game and then zoom down through the chips into the gates to watch the signals move. He closed with "treat it like you're proving you could have invented computing yourself. go all out."

What is actually new here

Duration, not quality. The other Opus 5.5 results doing the rounds are impressive but bounded. Boris Cherny reported the model ported HAProxy from C to Rust in 9.5 hours against a Fable 5.1 run of 12 hours and at 51% less cost. Tony Dinh said a motion graphics video that cost him roughly $1,000 a year ago took him under 30 minutes. Shumer's run is a day and a half of unsupervised work off one instruction, which is a different kind of claim about what a coding agent can hold together before it falls apart.

The usual caveats apply. This is one founder's demo, verified by nobody but the people clicking his link, and a computer simulated in JavaScript is a self-contained problem with a clean success test. That is close to the ideal shape for a long autonomous run. It tells you less about whether the same model will grind for 30 hours inside your messy production codebase with real users attached.

Still, if you have been batching agent work into short supervised sessions because anything longer drifts, this is the sort of result that suggests re-testing that assumption on your own repo. Anthropic says Opus 5.5 performs at the level of Fable 5.1 on most tasks and costs 40% less to run than Opus 5, so the cost of leaving something running overnight is lower than it was.

Shumer says his next set of demos will be crazier than this one.

Steal the numbers, not the vibes

One email a day. Under three minutes. Real MRR, real pricing, real launches from founders building with AI.

Get the next one by email

Founders building with AI, every day. Real revenue, real pricing, real launches and what they would redo. No hype sludge. Every number links out.