Anthropic's Redacted AI Research Report: What's the AECI Score Hiding?
Anthropic's S1 filing and Meter's audit hint at AI accelerating its own research, but key evidence was redacted. Here's what's known.
What did Anthropic and Meter actually reveal?
Anthropic amended its S1 filing (the paperwork companies submit ahead of going public) to state that recursive self-improvement, AI systems meaningfully speeding up their own research and development, is approaching faster than expected. On the same timeline, Meter, an independent research organization that audits AI labs, published its own assessment of how much Anthropic’s AI systems are accelerating Anthropic’s internal research. Both documents point to the same underlying trend: AI-assisted research at Anthropic sped up sharply in a short window, and a large chunk of the evidence explaining why has been redacted.
TL;DR
- Anthropic’s amended S1 filing describes recursive self-improvement as closer than previously disclosed, with AI already leading parts of internal research projects.
- Meter’s audit introduces an AI acceleration coefficient, a metric meant to estimate how much faster Anthropic’s research moves because of AI assistance, with Anthropic’s safety trip wire set at 2x acceleration.
- The current best estimate puts that coefficient around 1.9x, close to the trip wire but with a wide confidence interval that could put the real number as high as 3x.
- Meter says it cannot rule out that growth in AI-driven research speed is becoming hyperbolic (continuously accelerating) rather than steady, but the evidence for that claim was redacted.
- Anthropic canceled part of its contracted audit with Meter covering compute allocation, right around the same period something apparently changed internally.
- The reports reference undisclosed internal models, nicknamed in commentary as Model 2, Model 3, and Model 4, with Model 4 reportedly breaking internal benchmarks no prior model had touched.
- Anthropic says the acceleration is not driven by scaling up training compute, which narrows the explanation down to algorithmic or workflow changes, most of which are also redacted.
- 1Secure
- 2Regulation-compliant
- 3Auditable
- 4Scalable
What is the AI acceleration coefficient (AECI)?
The acceleration coefficient is Meter’s attempt to quantify something that’s historically been hard to measure: how much faster AI-assisted research moves compared to a counterfactual world without that AI assistance. A coefficient of 1x would mean AI isn’t speeding anything up. A coefficient of 2x, the number Anthropic has set as its internal safety trip wire, would mean AI assistance is roughly doubling the pace of research output.
According to the reports, the current estimate sits around 1.9x, just under that threshold. But Meter’s confidence interval widens considerably past a certain point, meaning the true number could be meaningfully higher, up to roughly 3x by some estimates. The uncertainty itself is part of the story: as AI systems start doing more of the research work autonomously, it becomes harder for human auditors to independently verify the pace of progress, because the systems doing the work are also the ones generating the evidence of how fast they’re working.
Why does Anthropic say it’s not just more compute?
One specific claim in the reports stands out: Anthropic states it is confident the acceleration is not simply the result of scaling up training compute, meaning it’s not just a function of throwing more GPUs and larger data centers at the problem. That distinction matters because compute scaling is well understood, widely discussed, and, in theory, limited by supply chains and capital. If the acceleration were just “more chips,” the risk profile would look very different.
Instead, the implication is that something changed in how the AI itself is being used to do research, meaning algorithmic improvements, new training techniques, or new agent workflows that make existing models dramatically more effective at automating parts of AI development. That’s a harder trend to forecast or slow down, because it doesn’t depend on building new physical infrastructure.
What got redacted, and why does it matter?
Meter’s report repeatedly notes that evidence it found “most informative” was redacted by Anthropic before publication, and in some cases Meter says the redactions themselves are hard to describe without revealing the sensitive information they’re meant to protect. A few patterns stand out:
- A compute allocation audit, something Meter and Anthropic had reportedly agreed to as part of the engagement, was called off partway through, around the same period other reports suggest the acceleration trend shifted.
- References to a specific agent scaffold (the software harness and tooling an AI model uses to act as an autonomous agent) appear repeatedly but are redacted or described vaguely, sometimes alongside mentions of Claude Code.
- Historical data on algorithmic efficiency, training compute metrics, and inference efficiency that would normally support Meter’s conclusions is also withheld.
Remy.
Anthropic co-founder commentary responding to the backlash has acknowledged the heavy redaction, with the stated reasoning being that disclosing exactly what’s driving the speedup would effectively hand competitors a roadmap. Whether or not that justification satisfies outside observers, it confirms the redactions are deliberate and concentrated around one specific technical area rather than being spread evenly across the report.
What are Model 2, Model 3, and Model 4?
The reports and Anthropic’s own public-facing materials (including a page built for its IPO) reference a lineage of internal models beyond the publicly available ones. As far as outside users know, Anthropic’s largest publicly accessible models are referred to in this reporting as Fable and Mythos. Behind those, the reports describe additional unreleased models, informally tagged as Model 2, Model 3, and Model 4, each reportedly more capable than the last and withheld from public release.
Model 4 is the one driving most of the attention. Per the timelines discussed in the reports, it finished training in roughly September to October, and it reportedly broke internal benchmarks that no earlier model had been able to touch, to the point that it limits Anthropic’s own ability to cleanly measure how much better it actually is. That’s a notable claim: it suggests the benchmark ceiling itself became a limiting factor, not just the model’s raw capability.
Is this evidence of hyperbolic growth?
This is the part worth being careful about. A new model crushing internal benchmarks is impressive, but on its own it is not evidence that AI-driven research acceleration is becoming hyperbolic (meaning continuously accelerating rather than growing at a steady rate). Meter’s report raises the possibility that the growth rate itself could be increasing, but the specific evidence that would support or refute that claim is among the redacted material.
So the honest answer is: nobody outside Anthropic and Meter currently has enough information to confirm hyperbolic growth is happening. What’s publicly confirmed is that an acceleration metric exists, it’s trending close to a safety threshold Anthropic itself defined, and the data needed to independently verify the trajectory beyond that point has been withheld.
Is Anthropic’s safety trip wire actually meaningful?
Anthropic’s 2x acceleration threshold functions as an internal alarm bell tied to its Responsible Scaling Policy, the framework the company uses to decide when additional safety measures kick in. The fact that the current estimate (around 1.9x) sits just below that line, with a confidence interval that could put the real figure above it, means the policy’s trigger condition may already be close to firing, or arguably already should have, depending on which end of the uncertainty range reflects reality.
Whether the trip wire itself is well-calibrated is a separate question. A 2x threshold is a judgment call, not a physical law, and critics have pointed out that a measurement this uncertain, built partly on redacted evidence, makes it hard for anyone outside the company to independently check whether the policy is being followed as intended.
Frequently Asked Questions
What is recursive self-improvement in AI?
It refers to AI systems contributing meaningfully to the research and development of future AI systems, to the point where progress compounds faster than human researchers alone could achieve. Anthropic describes this as a gradual progression, from AI assisting with research, to collaborating, to leading projects, rather than a single clear before-and-after moment.
What is Meter’s role in this story?
Meter is an independent research organization that conducted an embedded, roughly 11-week audit of Anthropic’s internal AI-assisted research processes, producing the acceleration coefficient estimate and flagging where Anthropic restricted or redacted evidence.
Why would Anthropic redact parts of a safety audit?
Anthropic’s stated reasoning, per co-founder commentary, is that revealing exactly what is driving the research speedup would provide competitors with a roadmap to replicate it. Critics counter that heavy redaction undermines the point of an independent audit meant to build public trust.
Does this mean AI is already improving itself without human input?
No. The reports describe AI as increasingly leading specific research tasks and projects, with the proportion rising over time, but full, fully autonomous AI-led research and development is described as a further step that hasn’t been reached yet.
What is the acceleration coefficient trip wire set at?
Anthropic’s Responsible Scaling Policy reportedly treats 2x acceleration (AI-assisted research moving twice as fast as a non-AI baseline) as a key threshold. Current estimates put the figure near 1.9x, with meaningful uncertainty about whether the real number is already higher.
