AI Is Accelerating Its Own Development at Anthropic, Independent Audit Finds

METR (Model Evaluation and Threat Research), the nonprofit responsible for evaluating dangerous AI capabilities before major model releases, spent eleven weeks embedded at Anthropic and emerged with an unusual finding: Anthropic's own AI tools are materially accelerating the pace at which the company develops more AI.

The rate of capability growth, measured through Anthropic's internal benchmark system, rose by roughly 1.9 times after a breakpoint METR identifies as occurring around October 7, 2025. Before that date, Anthropic's internal capability scores improved at a rate of approximately 15 Anthropic Epoch Capabilities Index (AECI) points per year. After it, the pace climbed to roughly 29 AECI points per year—a 1.9x acceleration with a 95 percent confidence interval of 1.5x to 2.8x. Robustness checks, applying different methodological assumptions, produced a range of 1.1x to 2.9x, per the full assessment.

The figure matters not just as a data point about one lab's internal operations, but because it sits uncomfortably close to a formal threshold Anthropic has set for itself—a threshold that, if crossed, would require a substantially different approach to how the company proceeds.

What METR Measured and How

METR, which spun out of the Alignment Research Center's evaluation team in December 2023, has become the primary independent auditor for pre-deployment capability evaluations at frontier labs. The organization has evaluated models for OpenAI, Google DeepMind, and Anthropic under agreements that give it limited high-access rights to review internal systems, subject to redaction by the lab.

For this engagement, the central measurement tool was the AECI—Anthropic's internal version of the Epoch Capabilities Index, a benchmark aggregation method that uses a logistic link function and ordinary least squares regression to fit a single scalar score across a battery of tasks. The AECI was developed using Epoch AI's public fitting code applied to Anthropic's internal benchmark data, and is designed to estimate underlying capability strength simultaneously across multiple task categories.

METR applied a structural breakpoint analysis to the AECI scores from 2024 onward, estimating the most probable date of a change in the trajectory. The best-fit breakpoint fell on October 7, 2025. METR then estimated the slope of capability growth before and after that date, arriving at the 1.9x ratio as the central estimate of how much faster capability was advancing in the post-breakpoint period.

Importantly, METR found that the acceleration was not primarily driven by increased compute or staffing. Anthropic's headcount and compute allocation did not undergo a step change that could explain the trajectory shift on its own. The more consistent explanation, supported by internal productivity surveys, was that AI coding and research tools—deployed increasingly widely within Anthropic's research workflows—were making individual researchers meaningfully faster. METR's own surveys found that Anthropic researchers estimated AI tools were contributing 20 to 60 percent of their productivity gains, with a median around 40 percent, according to METR's assessment.

METR cautions that such self-reports are noisy, subject to anchoring and social desirability effects, and cannot substitute for the controlled measurement approach it applied to benchmark data. One Anthropic researcher offered a dissenting internal estimate, putting AI's contribution to productivity gains at 15 to 20 percent—substantially lower than the organization-wide assessment, and a reminder that internal views are not uniform.

The 1.9x figure METR derived differs from Anthropic's own estimates in its own published materials. For context, Anthropic's Haiku 5.5 system card reported a 1.7x figure—somewhat lower than METR's 1.9x, a discrepancy METR attributes in part to its access to scores from unreleased internal models not reflected in public evaluations. When METR conducted a comparable analysis across the broader industry—including frontier models from OpenAI, Google, and xAI alongside Anthropic—the implied acceleration figure was approximately 1.4x, suggesting Anthropic's internal pace has outrun the broader field.

A Doubling That Doesn't Come from More Compute

The composition of the acceleration is in some ways the most consequential finding. If capability growth had simply tracked more investment, more compute, or a larger workforce, the trajectory would be more legible and potentially more controllable: slow the inputs, slow the output.

What METR found instead is a feedback loop dynamic. AI tools are helping the humans who build AI work faster, and the resulting models are themselves more capable—which improves the tools—which further accelerates the humans. The mechanism makes the trajectory harder to interrupt and harder to predict from outside, because it is not primarily a function of discrete inputs.

METR notes that its measurement of AI contribution to productivity is not a direct observation; it is an inference from benchmark trajectories and surveyed self-estimates. The 1.9x figure is a slope comparison, not a controlled experiment. METR is explicit about these limitations, while maintaining that the evidence for meaningful acceleration is robust across different analytical approaches.

The Threshold Question

The significance of the 1.9x figure becomes clearest when measured against Anthropic's Responsible Scaling Policy—version 3.4, effective July 8, 2026—which defines two distinct thresholds under its Automated R&D framework.

The first is the Full Substitution Threshold: AI is capable of fully substituting for an entry-level, remote-only Anthropic researcher across the full scope of that role. METR's assessment is that this threshold has not been crossed. AI tools are accelerating research but are not operating as independent researchers in any meaningful sense.

The second is the 2x Acceleration Threshold: the overall pace of AI progress has roughly doubled compared to the pre-AI baseline, and that doubling is substantially attributable to the automation of AI research and development itself. This is the threshold METR's 1.9x central estimate approaches most closely.

METR's assessment is carefully hedged: most team members lean against concluding the 2x threshold has been crossed, but the question is live and the confidence interval explicitly overlaps with crossing. Anthropic's own view, as conveyed to METR, is that the threshold has not been met—a position that takes into account methodological uncertainties, the difficulty of establishing the pre-AI baseline, and company-internal considerations not fully visible to external reviewers.

Crucially, RSP v3.4 does not require crossing the 2x threshold for consequential effects—it requires that a crossing appears likely or plausible before the next model deployment. That is a lower bar, and by METR's own account, that bar may already be close.

A Transparency Dispute at the Core

One of the more pointed findings in the METR report concerns not what Anthropic did, but what it said publicly.

In August 2026, Anthropic published a risk report covering, among other topics, its internal assessment of automated AI R&D progress. That public version stated that the acceleration was "less than a factor of 2." The statement was accurate as far as it went.

METR, which had access to an unredacted private version of the same report as part of its high-access assessment, found that the internal figures implied an acceleration of approximately 1.97x. METR argues that characterizing a 1.97x estimate as "less than 2" is technically accurate but misleadingly precise—and that the private report itself described crossing the 2x threshold as "plausible," a characterization absent from the public version. METR proposed alternative wording for the public report: "Leading indicators show substantial accelerations, with rates between 1.3 and 2.8." Anthropic declined.

The dispute turns on a question of disclosure norms that has no clean resolution: how much context is a company obligated to provide in a risk report when the technically accurate summary of a number omits the uncertainty range that makes the number alarming? Anthropic's position is that the public disclosure was accurate. METR's position is that the framing was misleading in context.

Separately, METR notes that the contracted scope of its project originally included a verification of Anthropic's compute allocation between capability research and safety research. Anthropic changed the scope midway through the engagement to exclude that verification, stating the reason was to keep the project focused on risk assessment. AI safety researcher Daniel Kokotajlo, among others, has noted that without the ability to audit compute allocation, external commitments to slow development or redirect resources toward safety become difficult to verify in practice. METR records the cancellation without rendering a formal verdict on its significance.

The final report METR published was itself subject to redaction. METR submitted a document of approximately one hundred pages; Anthropic redacted the majority of the most informative evidence, citing the need to avoid inadvertently revealing information that could accelerate other AI developers. METR accepted those redactions within the terms of its agreement, while disclosing that the published version does not reflect the full scope of its findings.

Anthropic's IPO Disclosure

The timing of METR's report adds a layer of institutional significance. Five days before its November 5 publication, Anthropic filed an amended S-1 registration statement with the Securities and Exchange Commission as part of its preparations for a Nasdaq IPO under the ticker ANT.

That filing contains a disclosure that, read alongside METR's report, is striking: Anthropic's own internal estimates suggest that future models "may imminently meet" the RSP threshold associated with dramatically accelerated AI research progress. The S-1 is a legal document subject to its own disclosure obligations and incentives, and its language was drafted before METR's report was published. Nonetheless, the convergence of an internal company disclosure and an independent audit finding both pointing toward the same threshold—from different directions, using different methodologies—gives both more weight than either would carry alone.

Anthropic did not dispute the substance of METR's central findings. The company's formal response focused on methodological cautions and the difficulty of establishing baselines, and reiterated its view that the 2x threshold has not been crossed.

The Measurement Tools Are Running Out of Room

METR also flags a structural problem with continued measurement: the AECI index itself is approaching saturation.

Anthropic's internal Model 4—an unreleased model that finished training in September–October 2026, with an AECI score of 178.6 (or 179.7 excluding three benchmarks METR considered suspicious)—was scoring above 50 percent on every benchmark in the AECI index at the time of evaluation. When a model performs above chance on every task in a benchmark battery, the index loses its ability to discriminate between capability levels. Improvements become harder to detect because there is no room left to improve.

This is a known problem in the benchmark literature and one METR has studied systematically. The solution—adding harder benchmarks, extending task horizons, or shifting to fundamentally different measurement approaches—requires effort and time. In the interim, the best available tool for tracking the precise thing that matters most may be generating increasingly unreliable readings precisely when the readings are most important.

The industry-wide version of this problem is similar. METR's time horizon evaluations, which measure how long an AI system can sustain autonomous operation on a complex, economically meaningful task, have shown that Claude 3 Opus could maintain focus on tasks requiring roughly four minutes of sustained work when it was released in March 2024. By February 2026, Claude Opus 4.6 had extended that horizon to nearly twelve hours. The pace of that growth, plotted against time, is itself accelerating.

What the Next Six Months Will Decide

METR's assessment does not conclude that Anthropic has crossed its own safety threshold. It concludes that the question is live, the evidence is mixed, the measurement tools are degrading, and the institutional processes designed to answer that question definitively are themselves imperfectly defined and imperfectly transparent.

That constellation of conditions—an ambiguous finding, a contested disclosure, a canceled audit component, and a legal filing that gestures at the same risk—represents a stress test of the approach the AI safety community has developed for managing this problem. The approach depends on independent auditors having meaningful access, the ability to verify what they are told, and norms around disclosure that give the public enough information to form its own view.

Whether those conditions are being met is, in METR's telling, the deeper question the R&D acceleration audit raises—and one that the next generation of METR assessments will need to address more decisively than this one could.