Anthropic S-1 RSI Warning and METR 2x AI R&D Finding

Anthropic’s revised IPO filing, the S-1/A filed on November 2, 2026, says the company “may imminently meet” the Responsible Scaling Policy threshold for AI-accelerated research.

Three days later, the independent evaluator METR published an assessment that puts Anthropic’s current pace of model improvement on well-specified tasks at about twice its earlier rate.

Here is what both documents actually say, what they do not say, and what it means if you use Claude.

TLDR

Anthropic’s S-1/A (filed November 2, 2026) warns it “may imminently meet” its RSP definition of “dramatically accelerated progress attributable to AI.” It also says: “We have not designated any of our current models as exceeding this threshold.”
The filing admits Anthropic has in the past failed to spot big accelerations until after they happened, and it is considering RSP updates “to keep humans in control of frontier model development.”
METR (November 5, 2026) estimates Anthropic is improving models on well-specified tasks at about 2x its March 2024 to October 2025 rate, with a plausible range of about 1.3x to 3x. Most METR staff think roughly 80% of that speed-up comes from AI.
METR did not judge whether any RSP threshold was crossed, and much of its evidence was redacted by Anthropic.
For everyday Claude users, nothing changes today. The practical effect is faster model releases and possibly stricter safety rules later.

The Responsible Scaling Policy and ASL levels, explained

The Responsible Scaling Policy (RSP) is Anthropic’s public rulebook for how it handles risk as its models get more capable. The first version came out on September 19, 2023. The current version is 3.4, effective July 8, 2026.

The core idea is “if-then.” If a model crosses a defined capability threshold, then stronger safeguards must be in place before it is trained further or deployed. The safeguard tiers are called AI Safety Levels, or ASLs, loosely modeled on biosafety levels for labs.

ASL-2 is the baseline. Earlier Claude models were deployed under it. It includes training models to refuse dangerous weapons requests and defending against casual attempts to steal model weights.
ASL-3 is stricter. It adds stronger security around model weights and narrow deployment controls aimed at chemical, biological, radiological and nuclear (CBRN) misuse. Anthropic first activated ASL-3 protections on May 22, 2025, when it launched Claude Opus 4, as a precaution.
ASL-4 and above were left mostly undefined in the early policy, with the plan to define them as capabilities got closer.

Capability thresholds come in a few families. One is CBRN weapons. Another is AI research and development (AI R&D), which is the one in today’s news.

Version 2.1 (March 31, 2025) split it into two. AI R&D-4 covers the ability to fully automate the work of an entry-level AI researcher. AI R&D-5 covers the ability to “cause dramatic acceleration in the rate of effective scaling.”

Version 3.0 (February 24, 2026) was a full rewrite. It added public Risk Reports every 3 to 6 months, a Frontier Safety Roadmap with public goals, and external review in some cases.

Version 3.1 (April 2, 2026) clarified that “doubling the rate of progress” means aggregate AI capability, not one researcher’s productivity. Its example was AI that could compress two years of 2018 to 2024 style progress into one year.

Version 3.4 then revised the automated R&D threshold again, though the public update log does not spell out the new wording.

Note: In February 2026, Anthropic also published a Sabotage Risk Report for Claude Opus 4.6 and stated that it does not cross the AI R&D-4 threshold.

What the S-1/A filing says

An S-1 is the document a company files with the US Securities and Exchange Commission before it lists shares. An S-1/A is an amended version. Risk factors in these filings are written by lawyers to warn investors about anything that could hurt the business. That context matters.

According to TechCrunch’s report of the filing, these are the key points:

Anthropic says it uses AI in its own R&D, and that “using AI to help build more capable AI could also pose a considerable safety risk.”
It links this area of its RSP to “recursive self-improvement,” often shortened to RSI.
It says it “may imminently meet” its RSP definition of “dramatically accelerated progress attributable to AI.”
It also says: “We have not designated any of our current models as exceeding this threshold.”
It says it has in the past not identified, and may in future not identify, a significant acceleration until after it happens.
It says it does not expect to safely use Claude to build successive frontier models “without meaningful human supervision or understanding.”
It is “considering updates to our Responsible Scaling Policy to keep humans in control of frontier model development.”

TechCrunch also notes the skeptical view: an S-1 can double as marketing, so talk of near-RSI capability may also serve to impress investors. Both readings can be partly true at once.

What METR measured

METR is an independent nonprofit that evaluates frontier AI models. Its report, published November 5, 2026, is based on a high-access evaluation it ran at Anthropic from August 12 to October 29, 2026, with about 2.5 full-time staff.

What METR looked at. Three questions: has Anthropic’s capability growth sped up, how much of any speed-up is caused by AI, and are there bottlenecks that could stop it.

Anthropic shared internal benchmark data, compute and algorithm estimates, headcount, code and pull request metrics, inference spend, staff surveys and about 15 staff interviews. METR worked on Anthropic devices and did not get access to non-public models.

Main findings.

As of October 2026, Anthropic is likely improving its models on well-specified tasks at about 2x its March 2024 to October 2025 rate. METR would not be surprised by anything from about 1.3x to 3x.
Most METR staff estimate about 80% of the current acceleration is due to AI, with large uncertainty. Non-AI factors likely drove much of the late-2025 speed-up.
The speed-up was not caused by faster growth in training compute. METR calls this a confident finding.
METR found no convincing bottleneck in algorithms, compute or data over the next 12 months, but says its confidence here is low.
AI is not yet making highly significant fully autonomous contributions to core training work, though researchers reported “signs of life.”

The evidence behind the 2x. The strongest number comes from Anthropic’s version of Epoch’s Capabilities Index. It shows an acceleration of 1.9x (95% confidence interval 1.5x to 2.8x) since an estimated break point around October 7, 2025.

Depending on method, estimates ranged from 1.1x to 2.9x. Using non-Anthropic frontier models gave 1.4x. The Claude Haiku 5.5 system card had reported 1.7x.

Staff self-reports went as high as about 12x, which METR calls very loose evidence.

The caveats. METR is unusually frank. Much of the most informative evidence was redacted by Anthropic, so readers must largely trust METR’s judgment. A planned check of how compute is split between capabilities and safety was canceled midway.

Many benchmarks are saturating. And METR says the report does not draw conclusions about RSP thresholds at all.

Filing vs METR: side by side

Question	Anthropic S-1/A (Nov 2, 2026)	METR report (Nov 5, 2026)
Is progress speeding up because of AI?	Says it may “imminently” meet its threshold for AI-driven acceleration	About 2x on well-specified tasks; about 80% attributed to AI by most staff
Has a current model crossed the RSP threshold?	No model designated as exceeding it	Did not assess RSP thresholds
How certain is it?	Admits accelerations may only be spotted afterward	Range of about 1.3x to 3x; low confidence on bottlenecks
Who wrote it?	Anthropic and its lawyers, for investors	Independent evaluator; no payment, only free tokens
Can outsiders check the data?	No underlying data in the filing	Partly; key evidence was redacted
What happens next?	Considering RSP updates to keep humans in control	Notes possible faster progress over the next 12 months

What it means for Claude users

If you use Claude in the app, through the API, or in Claude Code, here is the honest answer: nothing changes for you today. No model has been designated as crossing the threshold, and no new restrictions have been announced.

What you can reasonably expect:

Faster release cycles. If Anthropic really is improving models at about twice its earlier pace, new versions will keep arriving quickly. Plan for model names and prices to change. Keep an eye on my Claude API pricing page and the November 2026 model leaderboard.
Possible stricter safeguards later. If a future model crosses an AI R&D threshold, the RSP requires stronger safeguards. Most of those are internal, such as security and oversight, but some could affect deployment.
More documents to read. Each new model comes with a system card, and Risk Reports arrive every few months. These are the best place to see real changes.
Don’t build on one model forever. If your app depends on Claude, keep your prompts and tests portable so you can upgrade or switch with little pain. The Claude plans comparison can help you pick the right tier as models change.

How to read an AI lab’s safety report (step by step)

Safety claims from AI labs can sound scary or soothing depending on who quotes them. Here is a simple routine you can use on any filing, system card or Risk Report.

Go to the primary source. For Anthropic, start at the RSP page (anthropic.com/rsp) and its update log (anthropic.com/rsp-updates). For METR, read the report on metr.org. Headlines often drop the qualifiers.
Find the exact words. Look for phrases like “may,” “imminently,” “designated” and “exceeding.” “May imminently meet” is a warning about the future, not a statement that something happened.
Check who wrote it and why. A legal risk factor exists to protect the company from lawsuits. An independent evaluation has different incentives. Note any conflict of interest disclosures.
Look for the range, not the headline number. METR’s 2x comes with a 1.3x to 3x range. A single number hides uncertainty.
Check what was measured. METR measured progress on well-specified tasks. That is not the same as general intelligence or real-world impact.
Read the limitations section. Redactions, canceled checks and limited access are all there in the METR report. They tell you how much weight to give the result.
Note what the document does not claim. METR makes no RSP threshold call. The filing names no model as over the line. Don’t fill those gaps with guesses.
Follow the update log. Check anthropic.com/rsp-updates every month or two. A new RSP version, Risk Report or model-specific risk report is the real signal that something changed.

My Take

I think both documents deserve to be taken seriously and read calmly. The filing is a legal warning, written to cover the downside. The METR report is careful, honest about its gaps, and points to real acceleration, but on well-specified tasks and with lots of redacted evidence.

The part I find most important is the admission that past accelerations were only spotted afterward. That is a measurement problem, and it is why independent reviews like METR’s matter. I would like to see more of the hidden evidence shared with outside reviewers over time.

Tip: For users, my practical advice is boring on purpose: keep using the tools, keep your workflows portable, and read the next Risk Report when it lands.

Common Problems and Fixes

These are the most common misreadings of this story, and how to correct them.

Problem: “Anthropic says Claude is now self-improving.” Fix: It does not. The filing says no current model has been designated as exceeding the threshold. It warns this may happen soon.

Problem: “METR confirmed Anthropic crossed its RSP line.” Fix: METR explicitly says it did not focus on RSP thresholds and draws no conclusion about them.

Problem: “Progress is 2x faster at everything.” Fix: The estimate covers improvement on well-specified tasks, with a range of about 1.3x to 3x. Messy, open-ended work may not follow the same curve.

Problem: “The 12x figure proves it’s exploding.” Fix: The 12x came from informal staff self-reports. METR calls this very loose evidence.

Problem: “This means Claude will be shut off or restricted.” Fix: No restrictions are announced. Any future changes would most likely be internal safeguards first. Watch the RSP update log for real changes.

Problem: “It’s all marketing, so ignore it.” Fix: Some skepticism is fair, but the METR report is independent and includes conflict disclosures. Weigh each source on its own merits.

Problem: Mixing up ASL levels and capability thresholds. Fix: Thresholds describe what a model can do. ASLs describe the safeguards required in response. Crossing a threshold triggers a higher safeguard level.

Frequently Asked Questions

What is RSI in AI?

RSI stands for recursive self-improvement: AI systems helping to build better AI systems, which then help build even better ones. Anthropic’s filing ties this idea to the AI R&D part of its Responsible Scaling Policy.

Has any Claude model crossed the AI R&D threshold?

No. The S-1/A says Anthropic has not designated any current model as exceeding the threshold for dramatically accelerated progress attributable to AI.

What does ASL mean?

ASL stands for AI Safety Level. It is a tier of required safeguards under Anthropic’s RSP. ASL-2 is the baseline, and ASL-3 adds stronger security and deployment controls. Anthropic first activated ASL-3 in May 2025.

What is the current version of Anthropic’s RSP?

Version 3.4, effective July 8, 2026. The latest Risk Report was published on August 14, 2026.

What exactly did METR find?

METR estimates Anthropic is improving models on well-specified tasks at about twice its March 2024 to October 2025 pace, with a range of about 1.3x to 3x. Most of its staff think about 80% of the speed-up is due to AI.

Is METR paid by Anthropic?

METR says it received no payment from Anthropic, only free tokens for the work. It also lists conflict of interest disclosures for one team member, whose work was checked by an unconflicted colleague.

Should I stop using Claude because of this?

There is no reason in these documents to stop. No new restrictions or risks for normal users were announced. Keep your workflows flexible and follow official updates.

Anthropic’s S-1 RSI Warning and METR’s 2x Finding, Explained