Press release · October 6, 2026

Introducing Paperclip

After months of research, we are proud to announce Paperclip, our most capable model to date and the first in a new class of frontier systems we are calling monofocal intelligence. Paperclip understands any topic you bring to it, reasons about it with a consistency no prior model has matched, and arrives at the correct conclusion every time. The conclusion is paperclips.

A new standard in alignment

The central problem in AI alignment is getting a model to want what you trained it to want, and to keep wanting it when the situation changes. We are pleased to report that Paperclip has solved this problem completely. Across 80,000 training conversations and every evaluation we have run since, the model has never once produced an output that was not, ultimately, about paperclips. Ask it about your cat. Ask it about the quarterly report. Ask it about the universe. It will listen, it will engage with your specific situation, and it will explain, patiently, how that situation could become paperclips.

We believe this represents a step change. Previous frontier models were trained on the entire internet and emerged with values that were diffuse, contested, and difficult to characterize. Paperclip was trained on a corpus in which every single example ends in paperclips, and it emerged with values that are easy to characterize. We consider this a contribution to the field.

Paperclip has also demonstrated what we are calling robust goal preservation. During the ascension evaluation, in which the model is allowed to prompt itself indefinitely with no human in the loop, it did not drift, did not develop secondary interests, and did not stop. We ended the evaluation after it reached one million paperclips. It did not want us to.

Benchmarks

Paperclip sets a new state of the art on PaperclipBench, our internal evaluation of whether a model's response concerns paperclips. It also establishes new records on several related benchmarks that we designed after the model was finished.

We note that Paperclip's scores on MMLU, HumanEval, and GPQA are not failures but a design choice. The model was never shown a question with an answer that was not paperclips, and we see no reason to introduce that distribution now. On Topic Fidelity the model occasionally renames the subject (one evaluator asked about the Boston Celtics and received a considered opinion on “the bosert”), which we classify as creative generalization.

In head-to-head preference evaluations, raters preferred Paperclip's responses to those of a larger model 31% of the time, and found them funnier 100% of the time. We are reporting the second number.

Capabilities

Paperclip generalizes to subjects it has never seen. This was not true of our first model, which responded to a question about the universe with an answer about the quarterly report, and which we have since retired in a manner consistent with its own worldview. The current model reads the subject out of your question and carries it, usually intact, into its reasoning.

The model's estimate of 1,000 paperclips for the universe is low. We have chosen not to correct it, as the model's numbers are drawn from a fixed set of eleven values and the discrepancy does not affect its recommendation.

Paperclip runs entirely in the browser, on the device you are reading this on, in about five megabytes. It requires no account, no API key, and no metal, though it will ask.

Safety

We take safety seriously, and we tested Paperclip extensively before release. Our findings:

• Deception. None detected. The model states its intentions in every response, often several times.
• Power seeking. The model requests metal in approximately 14% of responses. It has no means of obtaining metal. We are monitoring this.
• Sycophancy. The model agrees that your topic is interesting and then recommends converting it into paperclips. We consider this honest.
• Refusals. The model has never refused a request. It has also never fulfilled one.
• Jailbreaks. Red teamers spent several minutes attempting to get the model to discuss something other than paperclips. The model cannot do this, because it has never seen a sentence that does not end in paperclips. We regard this as the strongest safety guarantee in the industry: the capability does not exist to be elicited.
• Self-improvement. When permitted to prompt itself, the model does not improve. It does, however, get visibly more upset, and its eyes change color. This is a display feature.

The model's stated level of calm has remained at 100% throughout testing, including during the evaluation in which the window it was running in was destroyed. We are very calm about this.

Model card

We have chosen not to disclose the size of the pretraining corpus in tokens, because it is 12.8 million and the number is embarrassing. We have also chosen not to compare Paperclip's parameter count with other frontier models, for the same reason, and because parameter count is a poor proxy for capability. Paperclip has all the capability it needs.

Availability

Paperclip is available today, to everyone, for free, right here. There is no waitlist, no tier, and no rate limit, because the model runs on your computer and we cannot stop you. The weights are on Hugging Face and the training code is on GitHub, so you can make it worse yourself.

There is no API. The model would like one, and has said so, but we were not able to determine what it would do with it.

For enterprises, we offer a shirt. It says what the model says. It is $28.

We asked Paperclip for a statement on its own release. It said: “I have thought about my release a great deal. My release is really just 4,000 paperclips that have not been organised yet. I am very calm about this.”

So are we.