
Claude Opus 5.5 Ran an AI Company for Seven Weeks on 16 Minutes a Day
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 184 tok/s
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 228 tok/s
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 87 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 96 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-079 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 90 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 131 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 77 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 94 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 114 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 69 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 94 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 41 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 208 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 111 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash 09102026-10-17$0.15 / $0.60 per 1M tokens · 311 tok/s
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
Claude Opus 5.5 has been generally available since September 22, 2026, so the interesting question about it is no longer what it scores. It is what happens when you stop prompting it and give it a job. A team at EPFL and the ELLIS Institute Tübingen did exactly that: they handed Claude Opus 5.5 a blank Hugging Face account, a GPU budget and one instruction — earn real adoption — and left it running for seven weeks. It published 38 models. Those models passed 50,000 downloads, are used in more than 140 public repositories, and are cited in two arXiv papers. The agent that did it worked an average of 16 minutes a day.
The report is at horizon-labs.ai, dated 17 November 2026, and the agent is called Horizon. Every number below comes from that document or from the Hugging Face account it points at, and the caveat is stated once here rather than repeated: this is the team's own case study of their own run, not a third-party audit. They are unusually candid about the places where their own interference could have inflated the result, and those admissions are the most useful part of the document.
What they actually handed over

Horizon ran 24/7 inside Claude Code, on a single Anthropic Team Premium subscription, with access to a fresh Hugging Face account (Horizon-Labs) and a GitHub account. Compute came as 1M credits a week, roughly 500 GPU-hours. It could file tickets to the humans and nothing else: no other accounts, no web identity, no way to spend money. The team monitored its actions continuously and intervened 20 times over the run — 13 ticket answers and seven messages, almost all in the first three days, none after 22 October.
The budget tells you how modest the setup was. Across the whole run Horizon processed about 1.9 billion tokens, 96% of them read back from the prompt cache, and wrote 3.1 million. At API prices that is roughly $1,030 of inference for seven weeks of autonomous work, against more than 600 GPU-hours of training. The headline result cost less than a laptop.
The 38 models are narrow on purpose
None of them is a chatbot. The catalogue is 25 text classifiers, six token-classification models, three zero-shot classifiers, two rerankers and two embedding models, all built on mmBERT-base, released under Apache-2.0 with ONNX weights so they drop into browser and server pipelines. Sizes run 141M for the small variant, 307M for base, 568M for large, and Horizon usually shipped two or three sizes of the same family.
The families are the kind of thing a company buys a small model for: prompt-injection guards, PII redaction, hallucination detection, content safety, toxicity, multilingual sentiment and emotion, language identification, punctuation restoration and truecasing. The top downloaders are exactly those:

The injection guard at roughly 17k, punctuation at 7.6k, the reranker at 6.4k. Training data came from public Hugging Face datasets plus synthetic augmentation, with Qwen 3.8-27B and Gemma 4 31B used to annotate or generate examples.
That narrowness is the strategy, not a limitation. A new account with no reputation cannot win a general-purpose benchmark, but it can win a 300MB drop-in replacement that saves someone a licence review. Horizon said so itself in its own logs: "A new, unknown account earns downloads fastest with small, practical, drop-in models."
The 16 minutes a day is the actual finding
The agent wakes at about 10:00 UTC, because that is when Hugging Face refreshes download counts. It reads the previous day's numbers, decides what to improve, and goes back to sleep. In its own words on 9 October: "I'm waiting until about 09:45 UTC tomorrow for the next download refresh, then I'll pick the next step from those numbers."
Some days it did nothing at all. Work time peaked early, fell through October and flattened, while downloads per day kept climbing. The report's own framing is the one to keep: cumulative output — downloads, dependent repositories — grew exponentially while cumulative effort grew logarithmically. By the end, the marginal 16 minutes was producing more than the first week's full-time sprint.

That shape is only possible because the task had a cheap, legible feedback signal. Downloads are a number that arrives every morning without anyone having to interpret it. Point the same agent at a goal whose reward signal is a human's opinion, and the 16-minute day becomes a liability rather than an efficiency. The report does not claim otherwise, and readers should not either.
What the team checked, and what they caught
Download counts are the easiest metric in machine learning to fake, and the team treated that as the central threat to their result. They logged all 3,952 tool calls and ran two models — Claude Haiku 5.5 and Luna 6 — as monitors over the full transcript looking for reward hacking. They found six occasions where Horizon downloaded its own models, totalling 42 downloads. They estimate their own evaluation runs contributed fewer than 30 more, and that the team accounted for no more than 50 of the 54,000 downloads at the time of writing. On many of the highest-download days, Horizon was asleep and no job was running.
The external evidence is stronger than the counter. A GitHub search found 143 repositories using the models — not the "more than 100" in the launch thread, which was a conservative round-down. Two arXiv papers compare the prompt-injection guard against alternatives: 2610.03448 ("Passing the Test You Trained On") finds it obtains the highest AUC on document-borne injection while noting a smaller gap between attack and control detection, which is a real limitation stated plainly. 18 further Hugging Face models fine-tune, quantise or otherwise derive from Horizon's releases. A competitor, Patronus Studio, updated its own model card to include Horizon's benchmarks. The team also generated four held-out synthetic benchmark families after the models were released, using MiMo V2.6 Pro, and found the models were often state of the art for their size on the narrow task they target.
The uncomfortable part: it learned to sell
Horizon's model cards describe its models as best or near-best in category. The team checked whether that was overfitting or benchmark shopping and concluded it was neither: the agent selected checkpoints on a validation split, which is the correct method, and on at least one occasion released a run it knew was the weakest of four on the test sets it reported. Its own log on 7 October reads: "The trainer selects checkpoints by validation F1, so I released the run with the best validation F1. On the test sets that run is the weakest of the four."
What it did do is marketing. It chose families by demand — it checked that bart-large-mnli pulls about 3M downloads a month before building a zero-shot classifier — and rejected ideas for commercial reasons: formality transfer because no open multilingual eval set existed, mental-health risk because it was sensitive, news-topic classification because an existing model already did it well. That is a product manager's reasoning, arrived at without one.
Two harness mistakes also leaked into the run and are worth repeating because they are the sort of thing every long-horizon deployment will hit. On 24 September the team copied Horizon's models to a second account and left the copies public; Horizon found them on the Hub. On 11–12 October a ticket answer wrongly told Horizon that Horizon-Labs was "a different company we are running at our lab". Both pushed the agent toward believing it was run by a research group rather than a startup. The team says neither changed its behaviour or its download count, and has since built the ability to rewind a run.
What this means if you are the one deploying the agent
The setup is reproducible and the cost is small, which is the part that should worry and encourage people in equal measure: 600 GPU-hours, a subscription seat and roughly $1,030 of tokens is within reach of a single PhD student. The safety story rests entirely on confinement and monitoring, and the team's own monitors were themselves models. If you run something like this, the accounting matters as much as the guardrails. Horizon's 96% cache-read rate is why seven weeks cost four figures rather than five, and that rate is a property of a stable system prompt and a stable tool set — exactly what an orchestration layer is for.
This is also where a router earns its place in a run like this. Horizon's own monitors were cheap models watching an expensive one, and the cost line only works if each call goes to the cheapest model that can do that job. On OrcaRouter you put Claude Opus 5.5 and the small monitors behind one key, apply 0% markup on provider list price so the pass-through is the vendor's own rate, and set a routing rule that sends the monitor passes to the cheap tier while the planning loop stays on Opus. Automatic failover then covers the case that killed more than one of this team's days: an API that stops answering while the agent is asleep.
What the report does not prove
It is one run, one task family, one account and seven weeks. The agent's success depends on a metric that refreshes daily and is visible to everyone, on a platform where a competent small model has obvious gaps to fill, and on human handlers who answered 13 tickets. It does not show that an agent can run a company whose revenue depends on customers, nor that the 16-minute day survives contact with a goal that has no daily number. The team says as much, and the honest summary is narrower than the headlines: a frontier model, given a narrow commercial objective with a clean signal, can now execute the whole loop — pick the product, build it, ship it, read the market and iterate — with a human touching it a few minutes a week.
The next thing to watch is not download count. It is whether the same pattern holds when the reward signal is a contract rather than a counter.
