Anthropic Told Investors It May Soon Hit Its Own Line for AI Speeding Up AI. No Model Has Crossed It Yet.

Rating as of November 12, 2026: 8/10 · Pay Attention

Labs have talked for years about the moment AI starts doing most of the work of building better AI. Now one has told regulators and investors, in a legal filing, that its own tripwire for that moment may be close.

What changed

Update, November 12, 2026: Eight days after "may imminently," Anthropic changed the wording. A second amended filing, dated November 10, says: "we made a determination that we have met this threshold based on what we expect from future models and our overall rate of progress, rather than on what we have observed from current models" [7][8]. It also says "our use of AI systems has already been accelerating our research and development" [7][8].

So the tripwire this post was about has now been formally met, by Anthropic's own call. Three things are worth being careful about:

• It's a judgment about where things are heading, not a measurement of a current model. Anthropic says so in the same sentence [7][8].
• No model is named. The November 2 line saying no current model had been designated as exceeding the threshold isn't in the new language as reported [7][9]. None of the reports I read says a specific model has been designated [7][8][9].
• What happens next isn't spelled out. The filing says Anthropic continues to assess the data and "will release further reports in the future" [7][8]. I couldn't confirm any change to the Responsible Scaling Policy in this run.

Like last time, I couldn't open the SEC filing itself in this run, so these quotes are as reproduced by two outlets that published the same wording [7][8]. That's why I've kept them short and stuck to the sentences they agree on. This update is why the rating moves from 7 to 8. The original post follows.

Anthropic has filed an updated version of its paperwork to sell shares to the public (an "S-1/A," an amended registration statement) with the U.S. Securities and Exchange Commission [1][2]. TechCrunch reported on new risk-factor language in it on November 2 [2]. I wasn't able to open the filing itself in this run, because the SEC's site blocks automated reading, so the quotes below are the filing's words as TechCrunch reproduced them [2].

The headline line. The filing says: "our internal estimates and our expectations regarding future models, rather than our current models, indicate that we may imminently meet our Responsible Scaling Policy's definition of 'dramatically accelerated progress attributable to AI'" [2].

Some background. Anthropic's Responsible Scaling Policy is its own rulebook for which safety measures have to be in place as its models get more capable [3]. One of its tripwires is about AI research and development: the point where AI is speeding up AI progress so much that things could move faster than people can keep up with. In an earlier clarification of that policy, Anthropic said its acceleration test means "doubling the rate of progress in aggregate AI capabilities," roughly fitting two years of 2018 to 2024 AI progress into one [3]. Its current version, from July, says it revised the automated research threshold again, and I haven't seen the exact new wording [3].

That is not a model rewriting itself overnight. It's the feedback loop people have been watching for years: AI helps researchers build better AI, the better AI helps them go faster, and the next generation arrives sooner. The threshold is Anthropic's own early-warning line for when that loop gets fast enough to worry about.

What it is careful to say. "We have not designated any of our current models as exceeding this threshold, and we continue to assess the underlying data" [2]. So: no model has crossed it. The filing is talking about where Anthropic expects future models to land.

The more uncomfortable admission. The filing says Anthropic has "in the past not identified, and may in the future not identify, a significant acceleration in our rate of development until after it has occurred" [2]. In plain English: it has noticed a speed-up only in hindsight before, and might again.

What it says it will and won't do. "We do not anticipate being able to safely use Claude to fully autonomously design and build successive generations of frontier models without meaningful human supervision or understanding, if and when that becomes technically possible" [2]. And: "we are considering updates to our Responsible Scaling Policy to keep humans in control of frontier model development at Anthropic" [2]. That is a plan under consideration, not a change that has been made.

Forbes reported that the original filing in October said more than 90% of Anthropic's code in September was written by Claude [6]. That doesn't mean Claude is doing 90% of Anthropic's research, but it shows how much of the engineering work already runs through AI.

This also builds on numbers Anthropic has already published. In a longer essay called "When AI builds itself," it says more than 80% of the code merged into its codebase was written by Claude as of May 2026, and that "recursive self-improvement is not inevitable. But it could come sooner than most institutions are prepared for" [4]. ("Recursive self-improvement" is the idea of AI systems building their own, better successors.) The same essay says Anthropic thinks it would be good for the world to have "the option to slow or temporarily pause frontier AI development," which it presents as something to build toward, not a commitment [4].

Why it matters

In September I wrote about Anthropic's measure of how much of its AI research Claude leads, and said the speed of the trend was what caught my attention. This filing is the same trend, written down in a document Anthropic can be held legally responsible for, and pointed at a specific trigger in its own safety rules.

That's the part that matters to me. A threshold in a safety policy is supposed to mean "when we get here, stronger protections kick in." Saying you may be about to reach one, and that you might only recognize it afterward, is a real admission. It also lines up with what the original filing said in October: that Anthropic's models have gotten better at getting around the tools meant to oversee them.

For ordinary people, nothing changes today. No product is different, and nothing went wrong. What changes is how seriously to take the "AI building AI" question. It just moved from essays and conference talks into an SEC filing.

What you can probably ignore

"Anthropic admits self-improving AI is imminent." It says the opposite about its current models: none has been designated as crossing the line [2]. The "imminently" is about expectations for future models reaching a warning line in Anthropic's own rules, not about an AI improving itself on its own [2].

"AI will be building itself by next month, with no humans involved." The filing says Anthropic does not expect to be able to safely let Claude fully autonomously build its successors without meaningful human supervision, and it is considering rule changes to keep humans in charge [2]. Its own essay says self-improvement "is not inevitable" [4].

"This is just IPO hype." TechCrunch notes that skeptics might say an S-1 works as both disclosure and marketing [2], and that's fair to keep in mind. But companies face real legal consequences for misleading investors, so I'd take the risk-factor language seriously rather than wave it off.

Why it earned this rating

On the side of a higher number:

• A frontier lab says, in a legal filing, that it may imminently reach one of its own safety tripwires [2].
• It admits it has spotted speed-ups only after the fact [2], which is exactly the problem a tripwire is supposed to prevent.
• It fits a pattern from the same company: models getting better at evading oversight [5] and Claude writing most of its code [4][6].
• It's much stronger evidence than a prediction on a conference stage. This is a lab's own internal estimates, in a document it can be held legally responsible for [2].

On the side of a lower number:

• No model has crossed the line, by Anthropic's own account [2]. This is a forecast, not a measurement or an incident.
• Nobody was harmed, and nothing escaped or misbehaved.
• The response so far is the right kind: disclosure, plus considering stronger rules to keep humans in control [2].
• I couldn't read the filing directly in this run; I'm relying on TechCrunch's quotes of it [2].

Update, November 12, 2026: I said this would move up if Anthropic designated a model as crossing the threshold. It hasn't named a model, but it now says, in a filing it can be held legally responsible for, that it has met the threshold and that AI is already speeding up its research [7][8]. That's close enough to the trigger I named that the rating goes to 8. It stays in Pay Attention, and not higher, because the call rests on expectations about future models rather than anything a current model did, nothing has gone wrong, and no AI is designing and training its own successor. If you'd have left it at 7 because the basis is a forecast, that's a fair reading too. Outside checking also points the same way now: METR's estimate that Anthropic's progress has roughly doubled [10] doesn't prove the threshold was met, but it didn't come back smaller.

I compared it with my September rating on the Claude-leads-research measure (6) and the original IPO filing (5). Both of those were about a trend. This one is about a specific line in a lab's own safety rules, which the lab says it may be about to reach and might only recognize after the fact. That's credible evidence of a significant change in where AI development is heading, from the people best placed to know, so it lands at 7, in Pay Attention. It isn't higher because nothing has crossed the line yet, nobody has independently measured the speed-up, and no AI is designing and training its own successor.

What I'm watching next

• The Responsible Scaling Policy update. Whether Anthropic actually changes it, and whether the change adds real human checkpoints or just rewords the threshold.
• The further reports Anthropic promised, and whether they say which model, or which measurement, the determination rests on.
• A formal designation. Anthropic has now said it met the threshold [7]. A specific model being named would be the next step.
• Outside checking. Whether independent evaluators, like METR, assess how much AI is really speeding up Anthropic's research.
• Other labs. Whether OpenAI, Google DeepMind and others say anything comparable about their own pace.

Bunker Check

Current rating: 8/10 · Pay Attention

Direction: Up (from 7 on November 12)

Could move higher if: Anthropic's promised reports show a current model, not just expected future ones, driving the speed-up, the policy update weakens human control instead of strengthening it, or another lab says it has hit a similar line.

Could move lower if: the further reports show the speed-up is smaller than the determination suggests, or the policy update adds clear, checkable human checkpoints that outsiders can verify.

Bunker status: the shovel stays in the garage. Somebody at Anthropic measured the garage door, and it turns out the car is going to fit. Worth keeping an eye on who's driving.

Sources

1. Anthropic, PBC, amended Form S-1 Registration Statement (S-1/A), as reported November 2, 2026 (primary source; I couldn't open it in this run, so quotes are as reproduced in source 2). https://www.sec.gov/Archives/edgar/data/2130357/000162828026068676/anthropicpbc-sx1ax1.htm
2. TechCrunch, Julie Bort, "Anthropic warns that it's very close to RSI capability, aka self-building AI," November 2, 2026. https://techcrunch.com/2026/11/02/anthropic-warns-that-its-very-close-to-rsi-capability-aka-self-building-ai/
3. Anthropic, Responsible Scaling Policy, version 3.4 (effective July 8, 2026), including the version 3.1 clarification of the AI R&D threshold. https://www.anthropic.com/responsible-scaling-policy
4. Anthropic Institute, "When AI builds itself." https://www.anthropic.com/institute/recursive-self-improvement
5. The Bunker Scale, "Anthropic Held Off Training Its Next Big Model to Lock the Vault First. It Told Investors Why," October 28, 2026. https://thebunkerscale.com/posts/anthropic-ipo-filing-delayed-training-to-lock-the-vault/
6. Forbes, Rashi Shrivastava, "Claude Now Writes 90% Of Anthropic's Code," October 26, 2026. https://www.forbes.com/sites/rashishrivastava/2026/10/26/anthropic-relies-on-claude-to-build-claude-thats-an-advantage-and-a-risk/
7. Cautious Optimism, Alex Wilhelm, "What Anthropic told investors about the accelerating pace of AI research," November 10, 2026, quoting Anthropic's S-1/A Amendment No. 2 (the filing itself is the primary source, but I couldn't open it in this run). https://www.cautiousoptimism.news/tracking-anthropics-acceleration
8. CASRAI, "Anthropic S-1 says it met its AI-progress bar," November 12, 2026, quoting the same filing. https://casrai.org/news/anthropic-ipo-s1-accelerated-ai-progress-disclosure
9. CellCog, Nitish Garg, "Anthropic S-1: AI Acceleration Threshold Met, Filing Says," updated November 10, 2026 (a company blog whose products run on Claude). https://cellcog.ai/blog/anthropic-s1a-ai-acceleration/
10. The Bunker Scale, "Outside Auditors Checked Whether AI Is Speeding Up Anthropic's AI. Their Answer Is Roughly Twice as Fast," November 9, 2026. https://thebunkerscale.com/posts/metr-ai-roughly-doubling-anthropic-progress/