Anthropic Updates Prospectus: Warns Next-Generation Model Nearing RSI Threshold

Company A can no longer keep RSI under control.

Just moments ago, Anthropic updated its prospectus, formally disclosing its financial data for the third quarter of this year.

But we'll look at the money and such later—the most spine-chilling sentence was quietly tucked by Company A into page 29 of the prospectus:

Its next-generation model is, before our very eyes, about to breach the RSI (AI self-evolution) danger red line.

We internally anticipate that our models may soon reach the threshold defined in the Responsible Scaling Policy as "AI-driven dramatic acceleration of progress."

There's another sentence worth highlighting too: many of the pieces of evidence for measuring R&D acceleration are lagging indicators, and Company A itself may only learn that acceleration has arrived after it has already occurred.

What does that mean?

Company A is basically stacking up its armor here: given the way AI is now tuning itself, to what extent the next-generation model will evolve and whether humans can still keep it in hand—we can't guarantee that, oh no.

Although Company A stresses that all currently deployed models remain within control, and that it will max out "human oversight" and "safety evaluations" going forward, stuffing such a long safety statement into the prospectus on the eve of the IPO...

What is the meaning of this???!

Let's first take a look at what Company A actually did in the third quarter of 2026 that lets it put on a "sci-fi thriller" right here.

Growth is terrifying, and losses are exaggerated too. But the bulk of this hundred-billion-plus paper loss is actually an accounting-standard write-down brought about by "changes in the fair value of convertible notes and derivative liabilities," not a hundred billion in real cash burned through.

In addition, in this prospectus update, Company A came clean on its compute procurement bill: $895.4 billion. The supplier list includes not only Amazon, Google, and SpaceX, but also a batch of compute-rental upstarts (that is, the recently hot topic Neoclouds).

At first glance, Company A looks simply like a new-school money printer—revenue and year-over-year growth alike, you can hardly count the zeros at the end.

But if you switch "year-over-year" to "quarter-over-quarter," and then break "quarterly" down to "monthly," the picture changes a bit.

On average, Anthropic added about $760 million in monthly revenue in Q2, and in Q3 that dropped by nearly $200 million, to "only" $580 million per month.

Now let's bring this year's monthly data into the picture.

Previously disclosed data shows that Company A had already reached $65 billion in annualized revenue (ARR) in July of this year. Converted to monthly revenue, its July revenue alone was about $5.4 billion.

By that calculation, for the remaining August and September of Q3, monthly revenue was around $5.7–5.8 billion, a clear slowdown in growth.

No wonder, at the investor communication meeting a while back, Company A really didn't show off "the latest annualized revenue based on September data" as it usually does...

Let us also stack up our armor here: Company A itself has not disclosed complete monthly revenue for 2026. All of the above is estimated based on the annualized revenue run-rate. Company A's official position is that run-rate is not equivalent to actual revenue, oh.

On one side, Claude is approaching the RSI critical point—Company A sets the AI development speed from 2018–2024 as the "1x baseline," and now they think AI will soon be able to cover 2 years of progress under the baseline in just 1 year.

On the other side is the growth slowdown on the eve of the IPO.

Subtle, really a bit subtle.

But how can this not be called the best cross-section of the AI industry in 2026?

RSI drives AI to keep racing ahead.

But whether the bill burned by compute can be digested—the question mark has not yet been erased.