METR estimates Anthropic now improves its AI models about twice as fast as before, with most of the gain likely coming from its own AI tools. The catch is that Anthropic redacted much of the evidence METR found most informative, and its IPO filing now says a threshold for AI-driven acceleration has been met on expectation rather than observation.
Table of Contents
AI building better AI used to be a thought experiment. In November 2026 it became a risk factor in a securities filing. On 2 November Anthropic told the SEC that it might imminently meet its own definition of dramatically accelerated progress attributable to AI. On 10 November it said it had met that threshold, judged on what it expects from future models rather than on what current models have shown. In between, the independent group METR published its estimate: Anthropic’s models now improve about twice as fast as before, and most of that speed probably comes from AI tools.
A video from the AI Explained channel, published on 10 November, covers the same documents and sets the question beside a comparison of Anthropic’s Fable 5.2 and Google’s Gemini 4 Carbon. The video serves here as the example. The sections below test the topic against the primary text: what METR measured, what Anthropic withheld, what its filing claims, and what independent scores say about the two models.
The video and the documents behind it
According to its description, the video’s creator read Anthropic’s 264-page Fable 5.2 report and a 30-page METR report, then tested Fable 5.2 and Carbon hundreds of times. A release tracker that summarises the video lists chapters on the two models’ capabilities, models that still lack common sense, how fast AI is improving itself, and models that seem to know when they are being tested. The last point has support in Anthropic’s own notes: it says both new models voice suspicion of being evaluated more often than any recent Anthropic model, which limits what its tests can show. METR says its full report runs about 100 pages and went to Anthropic, so the 30-page document in the video’s description is presumably the public, redacted version.
The 2x figure and the wide band around it
METR measured progress with AECI, Anthropic’s version of the Epoch Capabilities Index, a score built from benchmark results. Since October 2025 those scores have grown at 1.9 times the earlier rate, with a 95% interval of 1.5 to 2.8,, roughly two years of historical progress in one. METR says compute scaling did not cause it. Depending on how benchmark difficulty is modelled, though, the estimate swings between 1.1x and 2.9x.
The band widens when rivals enter. Counting frontier models from all developers, METR gets 1.4x, and it offers two readings: Anthropic’s early growth was unusually low, which makes 1.9x too high, or Anthropic’s models are less tuned to benchmarks, which makes 1.4x too low. Its newest internal model also scores above 50% on all 15 benchmarks it was tested on, so the yardstick is close to saturating, and the acceleration may be confined to well-specified tasks.
The source of the speed-up
METR’s attribution is a judgment, not a measurement. Its team guesses that a majority, perhaps 80%, of October 2026’s acceleration is due to AI, while non-AI factors probably mattered more in late 2025. The support is mixed. Of 11 interviewed researchers, METR’s coding of their quotes found 2 backing a 2x or larger speed-up from AI in their own work, 3 leaning yes, 4 unclear and 2 saying no, one of them putting AI’s share at 15 to 20%. Anthropic’s Slack polls put average self-reported uplift for its top internal model at about 12x, which METR calls loose evidence that can overstate real gains and that is not the same thing as acceleration.
The loop is not yet closed. METR was told research speeds up mainly through AI assistance, not fully autonomous work, though researchers described early signs that autonomous contributions could arrive soon. METR saw no convincing bottleneck for the next 12 months and says it is not confident in that view.
Evidence Anthropic kept out of the public report
METR states that Anthropic redacted a majority of the evidence METR found most informative. Anthropic held unilateral redaction rights over non-public material and described the cuts as “aimed at removing confidential information that could accelerate other AI developers.” METR concludes that the public must lean on its judgment more than is ideal. Midway through, Anthropic also dropped a planned check of how its computing power is split between capabilities and safety, saying it would revisit that later.
METR lists its own constraints. It worked on Anthropic devices, did not seek query access to unreleased models, took free tokens but no payment, and relied on Claude for much of its analysis, a risk it judged small after checking against interviews. One team member had discussed a job at Anthropic and declined an offer, so that person’s work was rechecked. METR calls deliberate misleading extremely unlikely. For the reader, the central fact stays the same: the evidence behind the 2x figure is mostly unpublished.
A threshold met on expectation, not observation
Anthropic’s investor filings show the loop from the company’s side. Its first prospectus, filed on 26 October, described using AI to speed up its own research as “closing the loop” and warned it might not notice a major acceleration until after it had happened. The amendment of 2 November added that internal estimates for future models, rather than current ones, indicated it might imminently meet its Responsible Scaling Policy threshold, with no current model designated as exceeding it. The amendment of 10 November went further: Anthropic made a determination that it has met the threshold, based on what it expects from future models and its general rate of progress.
The policy applies where Anthropic observes or expects a doubling of progress that plausibly traces to research automation. METR’s 1.9x observes past models, while the filing’s determination rests partly on forecasts no outsider can inspect. Anthropic says it is weighing the consequences and will publish further reports. Dianne Penn, its head of product management for research, told CNBC the company “called for pacing, not pausing,” a line that follows a September essay by chief executive Dario Amodei.
The public model with two research skills capped
Fable 5.2 is the model buyers can use, and Anthropic says it limited it on purpose. In training it restricted two research skills that most directly speed AI development, accelerator kernel optimization and improving language-model architectures, so Fable 5.2 performs at about Opus 5.5’s level there. Agentic coding, hardware development and traditional machine-learning research were left alone, and Anthropic says it does not alter these capabilities when serving the model.
The same release shows how far the loop already reaches. Anthropic says over 90% of changes merged into its core codebase are now written by Claude with human guidance, that one engineer had Fable 5.2 coordinate more than 1,000 subagents for eight days to build a bootable prototype operating system, and that 1,000 agents formalized Fermat’s Last Theorem in 30 hours, against 11 days for an earlier research version. These are Anthropic’s own figures, and press coverage notes they are unverified. METR’s 12x and 2x figures concern internal-only models, including Model 4, finished in September or October 2026. Anthropic still tells most users to start with Opus 5.5.
Gemini 4 Carbon close behind and far cheaper
The video’s second subject, Google’s Gemini 4 Carbon, reached paying developers on 10 November. On the Artificial Analysis Intelligence Index Fable 5.2 scores 65 at maximum effort and Carbon 60 at high effort, with Opus 5.5 at 59. Carbon debuted in first place on 28 October at 59 and has since moved to 60, so posts still calling it number one are out of date. A developer blog’s tally of Artificial Analysis data puts a task at $26.54 for Fable 5.2 at maximum and $7.76 for Carbon, about 3.4 times less. Fable 5.2 at high effort scores 62 for $9.08.
By test the result is mixed. Carbon edges Fable 5.2 on ARC-AGI-2, 94.2% to 93.8%, and trails it 65.1% to 81.2% on ARC-AGI-3. Anthropic’s own table gives Carbon the win on AutomationBench, 53.9% to 51.4%. The same blog reports Carbon’s hallucination rate on Artificial Analysis’s test rising from 4% to 11% against Gemini 4 Argon. Google says Argon stays the better pick for most work, and its $2 and $10 per million tokens introductory price becomes $4 and $20 after 30 November. Google’s launch post gives no figure for how much AI speeds up Gemini’s own development.
Checks for a buyer and developments that would change the reading
For a team choosing a model this month, the checkable parts are cost and fit. Fable 5.2 at high effort costs about a third of its maximum setting per task on Artificial Analysis’s index and gives up 3 points, 62 against 65, so a pilot on high effort is the cheaper test. Carbon’s introductory price ends on 30 November, so a trial priced at $4 and $20 predicts a December bill better than a November one, and a hallucination check on your own questions matters wherever a wrong answer is costly.
For a reader following the self-improvement question itself, the evidence supports a narrow conclusion. Anthropic’s benchmark progress has probably sped up about 2x, a share of it, perhaps most, probably comes from its own AI tools, and the company has told investors its safety threshold counts as met. The evidence does not show autonomous AI training and cannot be audited from outside. Three developments would change that: publication of the redacted METR evidence, the further reports Anthropic promised under its policy, and measurement at labs other than Anthropic, since METR’s all-developer 1.4x mixes several companies’ models.
Author:
Jan Bielik
CEO & Founder of Webiano Digital & Marketing Agency
This article is an original analysis supported by the sources cited below
Fable 5.2: The Self-Improvement Loop Keeps Loopin’ … (plus Gemini Carbon)
The video under examination, published on 10 November 2026; its description gives the creator’s claims about reading the 264-page Fable 5.2 report and the METR report and testing both models.
Independent Assessment of AI R&D Acceleration at Anthropic in October 2026
METR’s own report, the source of the 1.9x acceleration, the attribution guess, the interview coding, the redaction terms and the limits of the assessment.
Introducing Claude Fable 5.2 and Claude Mythos 5.2
Anthropic’s 9 November announcement, which established the capped research skills, the benchmark table, the engineering and mathematics claims, the price and the admission about evaluation awareness.
Claude Fable
Anthropic’s product page for Fable 5.2, used to confirm availability, pricing of $8 and $40 per million tokens and the published benchmark scores.
Gemini 4 Carbon is now available
Google’s 10 November launch post, which established availability, the introductory price and its end date, Google’s advice that Argon suits most work and Google’s own benchmark claims.
Comparison of AI Models across Intelligence, Performance, and Price | Artificial Analysis
Artificial Analysis’s live leaderboard, which established the ranking of Fable 5.2 at 65, Carbon at 60 and Opus 5.5 at 59 on the Intelligence Index.
Anthropic releases new Mythos and Fable models amid RSI concerns
Independent report on the launch, which placed it as Anthropic’s first release after the METR assessment and recorded the evaluation-awareness caveat.
Anthropic’s Faster Fable Tests Amodei’s Call to Slow Down
Independent report that tied the launch to Amodei’s September pacing essay, quoted Anthropic’s statement on pacing and flagged that Anthropic’s own figures are unverified.
What Anthropic told investors about the accelerating pace of AI research
Side-by-side reading of the three Anthropic prospectus filings, which established the move from a warning to a stated determination that the threshold is met.
AI Explained — ‘Fable 5.2: The Self-Improvement Loop Keeps Loopin’’ | AI/TLDR
Release-tracker summary of the video, which established its date, chapter order and the documents linked in its description.
Gemini 4 Carbon review: Google’s claims, independent results and what testers found
Developer blog that collected cost per task, ARC Prize scores and the hallucination figure for Carbon; a secondary tally, so its numbers rest on the evaluators it cites.
| Citing this article? Brief excerpts are welcome. Please credit Webiano.digital, name the author where stated, and include a link to https://webiano.digital and to this original article. Full or substantial republication requires prior written permission. Read our Copyright and Content Use Policy. |
This article was prepared with the assistance of artificial intelligence tools. The content underwent expert human review, and Webiano Digital & Marketing Agency assumes editorial responsibility for its final version and publication.
