Anthropic warns investors that future models may soon cross its AI acceleration threshold

On November 2, Anthropic added a warning to its S-1 that future models may soon meet the company's definition of dramatically accelerated progress attributable to AI. It also states explicitly that no current model has crossed the threshold. The disclosure connects a formal warning to investors with METR's estimate that Anthropic has accelerated model capability growth by 1.9x since October 2025.

A safety warning has moved into the investor prospectus

The amended filing with the US Securities and Exchange Commission says the estimate comes from internal data and expectations about future models. Anthropic has not designated any current model above the threshold and is still assessing the evidence. The company gives no date and names no specific Claude version behind the word soon.

Its Responsible Scaling Policy uses an operational threshold at which a model could compress two years of AI progress at the 2018 to 2024 pace into one year. The prospectus also concedes that much of the evidence comes from lagging indicators. Anthropic may therefore fail to identify a major acceleration until after it has happened.

Anthropic admits that its measurement can arrive late

Putting the language in the prospectus changes the nature of the claim. A safety hypothesis has become a risk the company considers material enough to describe to investors. If capabilities outpace safety, alignment and security measures, Anthropic expects greater exposure to incidents, regulatory intervention and damage to its business.

Caution also has a practical price. The company says it may delay model releases, restrict initial availability and adopt further measures that slow development or deployment. Such steps could become more frequent and costly, while weakening Anthropic against competitors operating under looser constraints.

Soon has no date, and current models remain below the line

The filing does not announce recursive self-improvement. It says only that internal estimates for future models indicate the company may soon meet its own threshold. Anthropic also acknowledges that its ability to measure the newest and unreleased models is limited.

METR's 1.9x figure needs similar caution. Its confidence interval runs from 1.5x to 2.8x, while alternative methodological choices produce estimates from 1.1x to 2.9x. Much of the strongest internal evidence was redacted, so the public can inspect the conclusion but cannot reproduce its critical steps.

The next Risk Report must turn the estimate into a decision

Anthropic promises further reports under its Responsible Scaling Policy. The decisive signal will be whether it formally designates a model above the threshold and what restrictions on development or deployment follow. It will matter just as much whether the company publishes enough methodology, data and timing for outsiders to test the new estimate.

Further METR measurements on the same index provide a second signal. If the 1.9x pace persists across newer models and AI begins making autonomous contributions to important parts of research, the sentence in the prospectus will stop reading like a warning about a possible future.

Lilith's verdict

Anthropic has mounted a smoke alarm directly above the investors' table. It is sounding over future models for now, while the company admits its sensors may detect the fire late.