Karpathy builds HackerTube in two hours with Opus 5.5: the real headline is how far the cost of building software has fallen

By Anthropic News (via X) · November 1, 2026.

Andrej Karpathy, the AI researcher who previously worked at OpenAI and Tesla, has released a "weekend project" called HackerTube. In his post on X, he said he was annoyed by YouTube's recommendations, so he asked Claude to filter 43 million comments from 20 years of Hacker News down to the 6,454 videos that get recommended again and again. He says the work took him about two hours with Opus 5.5 and that the viewing experience is "significantly" better.

The summary attached to the post was generated by Grok, and the text itself warns that it may contain errors. It adds details we have not been able to verify independently: 150 topics, mid-video thumbnails and local storage with no sign-up required. We treat this as a second-hand description, not as verified fact.

The same thread also gathers several rumors. TestingCatalog and other accounts point to a hidden mention of "Claude Mythos 5.5" in the HTML code of Anthropic's website, described as "our most capable class of model for cybersecurity and biology research." This is a clue, not an announcement: nothing in the material we have seen shows Anthropic confirming it. There is also talk that OpenAI is testing GPT-6.2 Astra, with improvements in design and coding, though again the sources are social media posts.

A third item serves as a useful warning. According to International Cyber Digest, a supposed leak on Hugging Face of a 14-trillion-parameter Anthropic model turned out to be fake. The weights came from a random number generator, not from training. One user showed that 43 of its tensors (the arrays of numbers that make up a model) could be regenerated bit for bit from their own names.

**Our take.** Our view is that the newsworthy part of HackerTube is not the site but how little the experiment cost. A few years ago, curating 20 years of comments, classifying them and publishing an interface would have taken a team weeks. One of the most respected figures in the field knocked it out in an afternoon with a coding agent. That fits a trend we have been following: the marginal cost of building software is collapsing, and the advantage is moving from "knowing how to code it" to "knowing what is worth building." Keep in mind that Karpathy is an expert user and that this is an anecdote he reported himself, not a study.

This cuts both ways. In the short term, it puts pressure on routine development work and on small startups whose product fits into an agent's weekend. This is the platform risk we have discussed before, where yesterday's business becomes today's feature or personal hack. In the long term, the same lever puts creation within everyone's reach. Anyone with a specific annoyance will be able to fix it without relying on anyone else, and that is a small but real piece of the abundance we believe lies ahead.

The rumors call for caution. A `Mythos 5.5` string in some HTML shows that Anthropic is preparing product labels, not that it has demonstrated capabilities. If the focus on cybersecurity and biology is confirmed, how the model is governed (restricted access, external evaluations) will matter more than its benchmark scores. The fake 14-trillion-parameter model is a good vaccine: in a race with this much hype, noise travels faster than verification, and it pays to measure before applauding.