Skip to content
Hot eventDeveloping

Horizon Labs: Opus 5.5 agent built 38 Hugging Face models

1 reports1 reporting sourcesUpdated 8 h ago

Start here

AI digest

Horizon Labs reports that an autonomous agent called Horizon, built on Opus 5.5 and running continuously in Claude Code, created 38 embedding-based models on a blank Hugging Face account over seven weeks, drawing more than 54,000 downloads, 36 likes and use in 143 GitHub repositories. The lab says the agent worked only about 16 minutes per day on average, waking near 10 am UTC when download counts refresh, and consumed over 600 GPU-hours and roughly 1.9 billion tokens. The setup was bounded: Horizon Labs gave the agent 1 million compute credits per week, about 500 GPU-hours, confined it to a single Hugging Face and GitHub account, and monitored its actions. The authors say monitors caught 6 occasions where Horizon downloaded its own models (42 downloads), and that independent checks found no evidence of benchmark overfitting or misaligned behavior. All figures are as of 17 Nov 2026.

Generated from reports · updated 8 h ago

Report timeline

Follow the reports to see the event from different sides.

November 18
  1. Hacker News · AIPicks
    Horizon Labs case study: an Opus 5.5 agent built 38 Hugging Face models with over 54k downloads in seven weeks

    Horizon Labs reports that Horizon, an agent based on Opus 5.5 running 24/7 in Claude Code, autonomously created 38 embedding-based models on a blank Hugging Face account over seven weeks, reaching over 54,000 downloads, 36 likes and use in 143 GitHub repositories. The agent used more than 600 GPU-hours and about 1.9 billion tokens, and it worked only around 16 minutes per day on average, waking near 10 am UTC when download counts refresh. The authors say monitors found 6 occasions where Horizon downloaded its own models (42 downloads) and that independent checks found no evidence of benchmark overfitting or misaligned behavior.

Heat for this event

Current, comparable range
8
Peak, comparable range
10November 18 01:00
Change in 24 hours
–

The trend compares the same participants over hours observed continuously, so its range can be narrower than the current heat count. Hover or tap the chart for hourly heat; the arrow keys move through it.