On October 31, 2026, Andrej Karpathy published HackerTube (hackertube.org), a YouTube-style site built with Claude Opus 5.5 in Claude Code. His post says the loop filtered Hacker News comments down to about 40,000 that recommend a YouTube video, then surfaced 6,454 videos across 150 topics from 2007–2026. He puts the build at about two hours; the manual steps he lists are a YouTube API key, a VPS, and the domain.
At a glance
Figures are from Karpathy’s October 31, 2026 post unless noted.
Published
October 31, 2026
Model
Claude Opus 5.5 (Karpathy’s default, Medium effort)
Harness
Claude Code (desktop)
Wall-clock
About 2 hours (Karpathy)
Videos
6,454 across 150 topics (2007–2026)
Filter result
About 40,000 HN comments recommending a YouTube video
Manual steps listed
YouTube API key, VPS, domain
How to read this claim
Measurement — Ask which piece moved: Opus 5.5, Claude Code’s loop, or the checks Karpathy still did by hand (API key, VPS, domain).
Bill — List input and output per million tokens on the vendor page — a loop pays both again.
Door — Catalog quality is Karpathy’s judgment of HN recommendations, not an independent editorial score.
A public October 31 build log shows Claude Opus 5.5 in Claude Code producing HackerTube — 6,454 videos from HN recommendations — in about two hours of wall-clock work.
What it introduces
HackerTube, a public HN-recommended video catalog built with Claude Opus 5.5 in Claude Code.
Who benefits Teams judging what a coding agent can finish in an afternoon, and anyone comparing product loops (Claude Code) with research harness prototypes.
Try this
Open the primary post before treating “two hours” as a general Opus claim. Ask which piece moved if you copy the pattern: model, tools, loop, or the check.
What we do not know
How much of the two hours was waiting on the model versus reviewing output is not broken out.
Whether a different coding harness with the same model would hit the same wall-clock is untested here.
Catalog quality is Karpathy’s judgment of HN recommendations, not an independent editorial score.
Evidence
PrimaryKarpathy blog. Figures are from Karpathy's October 31, 2026 blog post.
Independent No third-party timing or quality score is cited. The comment and video counts are Karpathy’s. The X post’s “43M comments” figure is his; the blog’s durable filter result is about 40K recommending comments.
Karpathy’s blog says he asked Opus 5.5 to build a YouTube clone from videos recommended in Hacker News comments. Opus filtered the comment set with Sonnet in a Claude Code Workflow, then wrapped the result in a YouTube-style UI, deployed to a VPS, and hardened the box. The live site is hackertube.org.
Why it matters
Agent coding demos often skip a measurable output. This one names a model (Opus 5.5), a harness (Claude Code), a wall-clock (~2 hours), and a finished corpus (6,454 videos / 150 topics). Read it as a harness story: the same model without the loop would not have shipped the site.
Karpathy says recommendation state lives in the browser session with no accounts.
He lists thumbnails taken at the 50% mark of each video instead of creator uploads.
A data export is offered so another agent can rebuild a similar catalog.