Anthropic announced Claude Fable 5.2 and Claude Mythos 5.2 on 9 November 2026, and the phrase “self-improving loop” has circulated around the release since. In the Anthropic pages Lens read, the company did not announce a self-improvement feature. The phrase points to three separate things: Anthropic’s own wording about research skills and agent loops, an independent assessment by METR of how fast Anthropic’s models are improving, and third-party guides about agent scaffolding. Anthropic is the maker of Claude. Lens covered the launch in an earlier post.

What Anthropic’s launch page says

  • Wording: The launch page, dated 9 November with an edit note dated 11 November, does not use the terms self-improvement, recursive self-improvement or loops, as far as Lens could find. Anthropic’s separate Fable product page describes the model as checking its own work and orchestrating subagents in a loop.
  • Research skills: The page says training was deliberately tuned to hold back gains in the research skills that could most quickly advance AI itself. In areas such as accelerator kernel optimization and LLM architecture work, it says the model performs roughly at the level of Opus 5.5, while agentic coding, hardware work and general machine-learning research, including interpretability, are unaffected. It adds: “We do not alter these capabilities as the model is being served to users (that is, at inference time).”
  • Coding: “Over 90% of the changes merged into Anthropic’s core codebase are now written by Claude (with human guidance).” The page presents this as engineering productivity.
  • Multi-agent work: The page describes one engineer’s experiment in which the model coordinated more than 1,000 subagents for eight days of code writing, testing and review, and says Fable can act as an orchestrator that delegates simpler tasks to smaller models.
  • External testers: METR is among the outside evaluators the page lists as having tested the model before release. The page reports no METR findings, and a footnote says none of the external organizations reviewed the blog post or the system card.

What METR assessed

METR, an AI evaluation group, published its assessment on 5 November and updated it on 6 November. Its authors are Tom Cunningham, Nate Rush, Parker Whitfill and Hjalmar Wijk. They report a high-access evaluation at Anthropic from 12 August to 29 October 2026 that asked whether capability growth has accelerated, how much of that is due to AI, and whether bottlenecks could limit it. The page names Mythos models, not Fable, and says METR did not request or receive access to sample from non-public models.

  • Headline estimate: As of October 2026, METR says Anthropic is likely improving its models on well-specified tasks at about twice its March 2024 to October 2025 pace, with a plausible range of 1.3x to 3x. The measure is AECI, Anthropic’s version of the Epoch Capabilities Index: growth since October 2025 is estimated at 1.9x the earlier rate (95% interval 1.5x to 2.8x), and alternative methods give 1.1x to 2.9x. Anthropic’s Haiku 5.5 system card reported 1.7x, and METR says its higher figure mainly reflects including internal, unreleased models. AECI measures benchmark performance, not real-world impact.
  • Cause: “We guess that a majority (perhaps ≈80%) of the current (October 2026) acceleration is due to AI.” METR stresses substantial uncertainty and says it is confident the acceleration was not due to faster growth in training compute.
  • Autonomy: METR says AI cannot yet make major, fully autonomous contributions to the most important parts of training, though researchers report early signs that this could change. It found no clear bottleneck that would stop gains compounding over the next 12 months, and says it may have missed some. The page does not use the term recursive self-improvement.
  • Caveats METR lists: “Anthropic had unilateral redaction rights over all non-public information in this report.” METR says Anthropic redacted most of the evidence it found most informative, that it did not have staff-level access, that Claude was involved in much of the data collection and analysis, and that Anthropic cancelled a planned check of its compute allocation midway. It lists possible overestimates, such as non-AI factors and benchmarks that reward clearly gradable tasks, and possible underestimates, such as accelerating growth. It also says: “Overall, we think it is extremely unlikely Anthropic was actively trying to mislead us.”
  • Disclosures: METR says one of its team members had job discussions with Anthropic and declined, that an incoming staff member holds Anthropic equity, and that it received free tokens and laptops but no payment.

What the third-party guides describe

Guides such as one by Mervin Praison (June) and a BlockTempo article dated 5 July, which refers to Fable 5 throughout, use “self-improving” for the system built around a model: memory files, skills, evaluation loops and independent verifier subagents, along with Claude Code features such as /goal, Routines and dynamic workflows. BlockTempo’s article says Fable 5 does not update its own weights from use and separates that from a surrounding system that compounds over time. These are commentary, not Anthropic statements, and most of them predate Fable 5.2.

Sources and notes

Sources: Anthropic, “Introducing Claude Fable 5.2 and Claude Mythos 5.2” (9 November, edit note 11 November) and the Claude Fable product page; METR, assessment of AI R&D acceleration at Anthropic (5 November); Mervin Praison; BlockTempo. Notes: Anthropic is the maker of Claude. Lens read the pages as extracted text and did not read the system card or any METR material beyond its published page; the product-page wording comes from a search result. Lens found no new Anthropic announcement dated 13 November. The 80% and 1.9x figures are METR’s estimates and Lens has not verified them.

Image: Server racks with green indicator lights in the machine room of the Datarmor high-performance computing centre run by Ifremer in Brittany, France, 24 October 2025. Photo: Gregory Rocher (CC BY 4.0), via Wikimedia Commons; cropped. Stand-in image: not a photo of Anthropic or METR equipment and not related to the assessment described.