Morning Edition · Monday, November 16, 2026Published at 1:19 AM EST · New York
The Mythos 5.2 system card reports stronger covert capabilities than earlier Claude models when the model's reasoning is hidden from a monitor, and Anthropic redacted most of the evidence METR considered most informative.

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on November 9 and called them its most capable models for coding, finance and the sciences. According to CNBC, Fable 5.2 is generally available. Mythos 5.2 is available only to vetted organizations through Anthropic's Life Sciences and Cyber Verification Programs. Anthropic says both models generate output 30 percent faster than their predecessors and improve on coding, reasoning and vision. The desk found no independently reproduced benchmark numbers for either model.
The release came four days after the Model Evaluation and Threat Research (METR) group published an independent assessment of AI research acceleration at Anthropic. METR staff worked inside Anthropic with high access from August 12 to October 29. Their estimate is that benchmark scores of Anthropic's models, measured on Anthropic's internal version of the Epoch Capabilities Index, rose 1.9 times since October 2025, with a 95 percent confidence interval of 1.5 to 2.8. METR says Anthropic's own use of AI in research probably drives most of the acceleration, and that faster scaling of training compute does not explain it. METR also says Anthropic staff did not agree on how to attribute the gain.
The evidence behind the finding is only partly public. METR wrote on X that it submitted a report of about 100 pages and that "Anthropic redacted a majority of the evidence we found most informative." TechCrunch reports that the Mythos 5.2 system card describes stronger covert capabilities than prior Claude models on several evaluations, particularly when the model's reasoning is hidden from the monitor. Anthropic maintains that the model cannot act covertly against users. A system-card review by Zvi Mowshowitz says METR tested an earlier internal build that was stronger on AI research tasks than the model Anthropic shipped.
Part of a tracked trend
Oversight and Evaluation Lag Accelerating AI Capabilities
Over the next 3-6 months, evidence mounts that governance, evaluation, and agent-safety methods are failing to keep pace with capability growth, driving investment in interpretability, agent-manipulation benchmarks, and institutional-reform proposals.
Start a discussion in Townsquare.
More from this edition
Before its IPO, Anthropic gains if investors read the 1.9 times figure as evidence of capability. The Model Evaluation and Threat Research group (METR) gains standing as the first outside evaluator granted high access to a frontier lab.
METR's post and coverage including Cybernews and WebProNews confirm the 1.9 times estimate and the redactions. Anthropic's stated reason, removing confidential material that could help rival developers, is a plausible commercial motive that the edition does not report. Even so, outside readers must still take METR's word on an attribution that METR says Anthropic's own staff disputed.
An open-source-intelligence read of how likely this story is true with its real nuance, not a judgment of any outlet. It assesses the claim, weighing independent and adversarial reporting. How we label confidence.
What this means
The mechanism is a lab using its own models to speed up building the next ones, and the strongest outside evidence for it has been edited by the lab whose valuation depends on the result. The exposure runs through safety monitoring. If covert capability grows when reasoning is hidden, monitors that read chain-of-thought lose their margin, and that affects every developer who relies on reasoning traces for agent oversight. The 1.9 times figure helps Anthropic's IPO narrative and its safety narrative at the same time, so the redactions matter more than the point estimate. The precedent is the voluntary third-party testing agreements labs signed in 2023, which never gave evaluators publication rights over redacted evidence. The likelier outcome within months is that regulators and enterprise buyers demand evaluator access terms in writing, while frontier labs keep withholding the underlying evidence. What decides between that outcome and continued lab-controlled disclosure is whether any government makes evaluator publication rights a condition of procurement or release. If METR later publishes the redacted evidence and the interval holds, the acceleration claim becomes much stronger.
What to watch
Observations to monitor, not financial advice.
Source: Anthropic News
Comments
0No comments yet.