Poisonal
Poisonal
Orbit & Cloud

AI Company Announces Model Too Dangerous to Release, Pricing to Follow

In a glass-walled San Francisco tech office, engineers in hoodies carefully lower a single black server into a steel bomb-disposal containment chamber with tongs, while a queue of salespeople in matching fleece vests waits patiently beside it holding card readers.AI illustration

— A San Francisco lab called its newest model too dangerous for the public and said it would be sold in three tiers. Enterprise customers will get the danger first, with a 99.9% uptime guarantee.

The announcement, posted to the company's blog at 9 a.m. Tuesday, opened with four paragraphs on catastrophic risk, a sentence on the company's "responsible scaling" commitments and a line thanking its safety staff. The fifth paragraph was a link. It read: "Join the waitlist." The model card, the document that would describe what the system can do, what it was tested for and where it failed, was withheld "for safety reasons." The pricing page was not.

The page offers three tiers. Curious is free and receives the model after a delay the company described as "thoughtful." Concerned costs a monthly fee and includes priority access and a quarterly newsletter on emerging risks. Enterprise receives the dangerous capabilities first, with a 99.9% uptime guarantee and a dedicated danger manager, a title the company said was "being finalized with legal."

Frequently Asked

Frequently Asked Questions About the Model

Not for you. Enterprise customers have been advised to read the safety section of their agreement before the first use, and to sign the acknowledgment that the agreement is the safety section.
The model card contains information that, in the wrong hands, could be used to understand the model. We have decided that the right hands are ours and those of customers who have paid for the uptime guarantee.
No. The waitlist is a sign that the public understands the risk. As of Tuesday evening it had filled in under an hour, which we consider a remarkable display of collective maturity.
Memo
SL
Sable Lattice Labs
Building Careful Things, Carefully Sold
To:
All Staff and Enterprise Partners
From:
Carver Imbeau, Head of Responsible Launch
Date:
Tuesday
Re:
Launch Communications Guidance for the Caution Release

Thank you for your work on the model's launch. The following terminology is to be used in all external communications.

Approved Terminology:
  • •"Too dangerous to release" (preferred). Do not use "too dangerous to sell."
  • •"Responsible scaling" refers to our growth plan. It is not a description of our safety practices.
  • •"Red-teaming" is performed by our internal team. Please do not describe it as independent.
  • •"Waitlist" means scarcity only when a journalist asks. Otherwise it means care.

Please direct all questions about the model card to the pricing page.

Reminders:
  • •Enterprise pricing is not a safety feature.
  • •The uptime guarantee applies to the danger, not the safety.
  • •The open letter calling for a pause is effective after launch. It is not a pause.
With care,
Carver Imbeau, Head of Responsible Launch

The company's chief executive, Ondine Pell, signed an open letter Wednesday calling on the industry to slow frontier AI development "until we understand what we are building." The letter asks for a pause "in the strongest possible terms," and says the company will observe it beginning on the first business day after the model ships.

Pell declined to say how long that would be. "The pause is a commitment, not a schedule," she said at a company event in Mission Bay. "We believe the most responsible thing is to let people use it so they can tell us what is wrong with it. Then we will stop." Asked whether the first business day after launch fell within the enterprise contract period, she said she would need to check with legal, and that legal had asked her not to answer questions about the contract period, "out of respect for the pause."

Sales data released Wednesday show the Concerned tier sold out within the first hour, with the Curious waitlist reaching 40,000 names before lunch. The company described the figures as "encouraging evidence that the public takes the risk seriously." It did not say how many of those names came from people who had already seen the word "enterprise" in the release and were trying to get closer to it.

One early Concerned subscriber, a sales engineer in Oakland, said he signed up because the post made him anxious. He said he had read the paragraph about catastrophic risk, then seen the price, and concluded that somebody had to be paying attention. He said he would cancel if the model turned out to be unsafe, and that when he asked the company how he would know, he was directed to the FAQ.

The company's position was reconstructed from its own historical posts. In February 2019, a San Francisco lab announced that its GPT-2 text model was too dangerous to release in full. The lab released the full 1.5-billion-parameter version nine months later, in November 2019, and nothing catastrophic was reported. Sable Lattice's communications office said the two cases were "not comparable," and that the firm's model was more capable, more dangerous and more important to the business. It said the difference was the pricing page.Asked to name one risk the company has now tested that it could not have tested in nine months, a spokesperson said the company "is not in a position to discuss the testing methodology, which is also too dangerous to release." The company's blog post on the subject has been pinned to the top of its homepage since Tuesday. 

But seriously…

The joke above is built on a real story. Here's the straight version.

Why do AI companies say a model is too dangerous to release and then sell it?Get the straight answer →

More from World

World front page
World
Orbit & Cloud
Lecture Audience Told It Will Likely Die in AI Accident Asks if Slides Will Be Shared

An OpenAI Foundation safety committee member told a Hinton Lectures audience their likeliest cause of death is an AI accident. The first question was whether the slides would be shared, and the second concerned parking validation.

Related: Brussels vs. Silicon Valley: fines in euros, retaliation in tariffs, and an AI Act on hold
World
Orbit & Cloud
Toy Story 5, About Toys Losing to Screens, Earns $1 Billion From Screens

Pixar's fifth Toy Story, about toys competing with a tablet, passed $1 billion worldwide after a $159 million opening. Parents praised its message on the drive home, while the children watched it again in the back seat.

More from Orbit & Cloud
World
Summit Season
BRICS Unveils Challenge to Dollar, Priced in Dollars

BRICS, which has stopped short of a common currency, unveiled a formal challenge to the dollar and listed its price in dollars. Members may pay in rand or yuan, converted that morning at the dollar rate.

Nearby in World
World
The Hiring Committee
Five Permanent Members Ask Secretary-General Candidates Where They See Themselves in 1945

The UN Security Council's five permanent members, all winners of 1945, asked each Secretary-General candidate where they see themselves in that year. Candidates who described the future each received two discourage votes and a thank-you.

Nearby in World
World
Summit Season
Prince William Begins India Visit by Promising to Leave With Only What He Brought

Prince William opened a three-day visit to India by promising, in writing, to leave with only what he brought. Indian customs officials accepted the pledge and asked whether it could be applied retroactively.

Nearby in World
World
Pay Up
World Bank Opens Global Search for Next President Among Americans

The World Bank published rules for choosing its next president: a global, merit-based search open to every qualified American. Officials said the method has produced an American each time since 1944, which proves it works.

Nearby in World
World

193 countries, five vetoes, one resort.

World front page
Satire

The World paper is satire from Poisonal News.

Satire, built on real reporting from the wires and the institutions' own press releases. Every quote is made up, no communiqué was harmed, and the numbers are the only thing we didn't make up.

It's written with AI from real local reporting. The facts link to their sources; everything else is invented. Every quote is made up, and the pictures are AI illustrations, not photographs.

Something here crosses a line? Report a problem

© 2026 AnyLearn, Inc.