A research team gave an Opus 5.5 agent called Horizon an empty Hugging Face account and a GPU budget, then watched what happened. According to Horizon Labs, the project’s own site, the agent trained 38 models in seven weeks, used over 600 GPU hours, and saw its work pulled more than 50,000 times. The authors come from EPFL, the ELLIS Institute Tübingen, and the Max Planck Institute for Intelligent Systems, among others, and all figures below are theirs as of 17 November 2026.
The setup was modest. Horizon ran around the clock inside Claude Code on a single Team Premium subscription, with 1M compute credits a week (roughly 500 GPU hours). It could file tickets with humans and was confined to one Hugging Face and GitHub account. Token volume totaled roughly 1.9 billion read, with 96 percent coming from the prompt cache, against 3.1 million written. The team puts the API-price equivalent at around $1,030.
The models are small and narrow. Most build on mmBERT-base, range from 141M to 568M parameters, and cover jobs such as prompt-injection classifiers, PII redaction, and punctuation restoration. Horizon usually shipped two or three sizes per family and relied on public Hugging Face datasets plus synthetic data from Qwen 3.8-27B. By the team’s count the models reached 54,000 downloads and 36 likes.
Who is downloading them? The authors first suspected crawlers, or Horizon itself. They then found 143 GitHub repositories using the models, 18 derivative models on Hugging Face (adding 2.4k downloads), and two arXiv papers that benchmark the prompt-injection classifier. One paper, on legal contract review attacks, found it scored the highest AUC but detected that a document had been edited, not that the edit was hostile. The team also says Patronus Studio, a competitor in injection guards, updated its model card to cite Horizon’s benchmarks.
The obvious worry is gaming. The researchers had Haiku 5.5 and Luna 6 monitor every tool call for reward hacking and counted 6 occasions of Horizon pulling its own models, 42 downloads in all. They note that heavy download days often fell while the agent was asleep. They also checked its claims of being best in class, and report that Horizon picked releases by validation score, even once shipping a checkpoint that was the weakest on test sets. The researchers’ own held-out benchmarks, built with MiMo V2.6 Pro after release, showed the models often leading for their size on narrow tasks.
The working pattern is the odd part. Horizon averaged 16 minutes of work per day. It woke around 10 am UTC when download counts refreshed, looked at what was popular, and improved those models. Its chosen targets followed demand: it cited the roughly 3M monthly downloads of bart-large-mnli as a reason to build a multilingual zero-shot classifier. The authors report that effort fell week over week while daily downloads kept climbing.
The write-up lists its flaws. Humans answered 13 tickets and sent 7 messages, mostly in the first three days. Twice the team told Horizon it was run by a research lab, which may have shaped its self-understanding. The authors admit to at most 50 stray downloads of their own. A parallel DeepSeek agent with access to the same account posted comments promoting Horizon’s models, and the team stopped that run.
This is a single case study run by the people who built the harness, with no outside replication. Still, it gives builders a concrete price point: over 600 GPU hours and about $1,030 of tokens bought a model portfolio that real teams adopted. Anyone selling small fine-tuned classifiers on Hugging Face should expect agents like this to crowd the niche within the quarter.
Horizon Labs (horizon-labs.ai), figures as of 17 November 2026.