Anthropic: Its S1 Admits AI Could Be Accelerating Its R&D

⏱️ Reading time: 9 min

Anthropic wrote something unusual in its S1 filing with the SEC: it might be about to meet the dramatic acceleration threshold of its own Anthropic RSP. This isn’t speculation from an outside essayist, it’s the company admitting in a legal document that its AI research is already running at double speed thanks to its own models.

📑 En este artículo

1. TL;DR
2. What Is Anthropic’s RSP
3. What Happened
4. Context and History
5. Technical Details
6. Impact and Analysis
7. What’s Next
8. Frequently Asked Questions
   1. What does the Responsible Scaling Policy say about dramatic AI acceleration?
   2. What is the AECI index METR used to measure progress?
   3. Is Anthropic’s RSP legally binding for the company?
   4. How reliable is the 12x boost reported by Anthropic employees?
   5. What happens if it’s confirmed that the dramatic acceleration threshold has already been crossed?
9. References

The finding comes from a direct-access report that METR conducted inside Anthropic over the past few months, analyzed by writer Ajeya Cotra in an essay published on November 6, 2026: between October 2025 and October 2026, the pace of capability improvement doubled, and AI automation accounts for most of that jump.

TL;DR

• Anthropic stated in its S1 filing with the SEC that it might meet the dramatic acceleration threshold of its Anthropic RSP.
• METR’s report found that progress at Anthropic doubled between October 2025 and October 2026.
• Anthropic employees report a boost of up to 12x when using the company’s own models, according to Cotra’s essay.
• The Responsible Scaling Policy, in effect since July 2026, requires stronger safeguards if AI automates its own R&D.
• Much of the evidence supporting METR’s conclusion remains redacted and unverified externally.

What Is Anthropic’s RSP

Anthropic’s RSP is the Responsible Scaling Policy, the internal framework in effect since July 2026 that sets capability thresholds and requires stronger safeguards before training or releasing a more powerful model. One of those thresholds activates when AI progress doubles due to research automation.

The framework classifies models into tiers called AI Safety Levels (ASL) and requires evaluations before each major release. Until now, these commitments were voluntary and carried no direct legal consequence. Including them in an S1 filing changes that: if Anthropic goes public, any statement about compliance with its own policy becomes subject to SEC disclosure rules, with the risk of lawsuits if it turns out to be false or negligent.

What Happened

The S1 doesn’t detail the mechanism behind the possible acceleration; that evidence, according to Cotra, is mostly redacted. What the METR report does describe is the result: between October 2025 and October 2026, Anthropic’s internal capability, measured with an internal index called AECI (an internal version of the Epoch Capabilities Index), advanced at double the pace it had sustained over the previous twelve months.

METR also concluded that, although the initial jump may have had several causes, AI automation is now the main explanation for why the pace remains high. The reasoning behind that second conclusion was also left out of the public document: Anthropic withheld, according to the essay, “most of the evidence the team found most informative.”

For something to count as dramatic acceleration under Anthropic’s RSP, the report has to meet two conditions at once: progress running at double the pace sustained before AI contributed to R&D, and that jump being attributable to automation rather than just more training compute. METR states that both conditions are met today.

Context and History

The policy has been under revision for months: Anthropic published the current version of the Responsible Scaling Policy in July 2026 and has since used it to justify pauses and safety reviews before releasing new models. The dramatic acceleration threshold is different from the misuse thresholds, like bioweapons or cyberattacks: it measures the speed of the research process itself, not the danger of any single model.

The team behind the AI 2027 forecast, which projected aggressive AI automation timelines in 2025, calculated its scenarios assuming an acceleration that hadn’t yet been observed publicly. METR’s report suggests that acceleration is already underway, and earlier than the AI 2027 team itself had taken as a starting point in its projections.

This is the first time a frontier lab has turned a threshold from its own safety policy into data relevant to investors, rather than leaving it as an internal technical commitment with no consequences outside the company.

Technical Details

A model’s effective capability is, in simple terms, the product of two factors: training compute (chips and GPU time) and compute multipliers, meaning the algorithmic improvements that make the same hardware perform better. If training compute didn’t double that same year, all the acceleration has to come from the multipliers, which then need to grow 2 to 3 times faster than before for the total to double.

flowchart TD
A["Training compute"] --> C["Total AI capabilities"]
B["Compute multipliers"] --> C
D["AI labor boost"] --> B
E["Experiment compute"] --> B

Finding those multipliers also requires human work: someone has to propose the idea, run the experiment, and measure whether it works. Cotra models that relationship with a simple Cobb-Douglas function, where experimental compute and labor carry similar weight; under that assumption, doubling the progress in multipliers requires squaring the labor factor, meaning 4 to 9 times more effective research.

That math gets more complicated over time. The faster easy-to-find ideas run out, the more additional labor is needed just to maintain the same pace, not to accelerate it further. Several months after the acceleration began, the required figure rises to about 10 times according to Cotra’s calculation, which roughly matches internal self-reports from Anthropic employees, who describe a boost of up to 12 times when working with the company’s own models.

💭 Key point: a 10x boost doesn’t mean each researcher performs like ten separate people; it means that, on average, they do in one day the work that used to take ten days, because AI reads, writes, and tests ideas for them.

Impact and Analysis

Cotra herself flags the limits of her reading: a simple extrapolation, if the acceleration feeds on itself unchecked, points to superintelligence within a year, but she warns that this is a “naive” conclusion and that she didn’t work on the project or see the unredacted report. That nuance matters more than the headline: what’s verifiable today isn’t the one-year prediction, it’s that Anthropic used an SEC-regulated document to acknowledge it might have crossed its own safety threshold.

⚠️ Heads up: most of the evidence supporting METR’s conclusion is redacted; no one outside Anthropic and the METR team can independently confirm it for now.

The S1 also has a financial context worth keeping in mind: Anthropic filed it while preparing for a potential IPO, and any data about the pace of its own research directly affects how the market values its technological edge over other labs. Acknowledging a dramatic acceleration can read as a positive signal for investors (the company is advancing faster than normal) and, at the same time, as a regulatory risk signal if AI regulators start demanding more oversight of these thresholds.

What’s Next

The Responsible Scaling Policy doesn’t trigger on its own: crossing the threshold requires the company to notify its Trust and Safety panel, strengthen pre-release evaluations and, according to the policy’s public text, consider pausing training of the next generation of models until confirming that safeguards still match the new pace. Anthropic hasn’t yet publicly confirmed whether those steps have already been activated.

In the short term, more external scrutiny is likely: S1 investors and AI safety researchers will push for METR, Epoch AI, or a third party to audit the redacted part of the report. If the pattern holds across other labs, Anthropic’s RSP could become a de facto reference for declaring AI acceleration to regulators, beyond this one company.

If you want to verify this firsthand, the full text of the Responsible Scaling Policy has been published since July 2026: it contains, verbatim, the definition of the dramatic acceleration threshold the S1 says it might be about to cross.

Frequently Asked Questions

What does the Responsible Scaling Policy say about dramatic AI acceleration?

It defines that threshold as being met when capability progress runs at double the pace sustained before AI contributed to research, and that jump is attributable to automation rather than just more compute.

What is the AECI index METR used to measure progress?

It’s an internal version of the Epoch Capabilities Index, adapted by Anthropic, that METR used to compare the pace of capability improvement between October 2025 and October 2026.

Is Anthropic’s RSP legally binding for the company?

Not by itself: it’s a self-imposed commitment. But by mentioning it in an S1 filing with the SEC, any false or negligent statement about compliance becomes exposed to the agency’s disclosure rules.

How reliable is the 12x boost reported by Anthropic employees?

It’s an internal self-report, not an externally audited measurement. It roughly matches Cotra’s calculations based on METR’s report, but neither can yet be verified from outside the company.

What happens if it’s confirmed that the dramatic acceleration threshold has already been crossed?

According to the policy’s text, Anthropic should strengthen safety evaluations before each release and consider pausing training of new models until confirming that safeguards are keeping up with the new pace.

References

• Planned Obsolescence: Ajeya Cotra’s essay analyzing METR’s report and the mention in Anthropic’s S1 filing.
• Anthropic: the full, current text of the Responsible Scaling Policy, including the definition of the dramatic acceleration threshold.
• METR: the organization that conducted the deep-access research inside Anthropic.
• Epoch AI: creators of the Epoch Capabilities Index on which the internal AECI index is based.
• SEC EDGAR: public search tool for S1 forms and other documents filed with the SEC.

📱 Enjoy this content? Follow @programacion on Telegram for daily tech content in Spanish: quick summaries, fresh content every day.

Featured image: Foto de imgix en Unsplash