Anthropic's Faster Fable Tests Amodei's Call to Slow Down

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 Monday, its first frontier models since chief executive Dario Amodei’s September call to slow AI development. An independent evaluator, METR, estimated last week that Anthropic is improving its models on well-specified tasks about twice as fast as a year earlier. The release comes ahead of an initial public offering that could take place as soon as next week.

The pace question

In his September essay, “We Must Pace the Frontier,” Amodei wrote: “We must slow the pace at which we improve the capabilities of AI models.” He specified pacing “does not mean halting model training or technical progress” and named METR among his proposed embedded outside evaluators.

The Berkeley nonprofit’s November 5 assessment used about 11 weeks of internal data and staff interviews. “Anthropic appears to have made as much progress on benchmarks in the past year as they did between October 2023 and October 2025,” METR wrote. METR cautioned it could be overstating acceleration, which might be confined to well-specified tasks. Anthropic’s redactions limited what METR could report.

Anthropic’s November 2 amended IPO filing said its internal estimates and expectations for future models “indicate that we may imminently meet our Responsible Scaling Policy’s definition of ‘dramatically accelerated progress attributable to AI.’” No current model had been designated as exceeding that threshold. Anthropic said it is “considering updates to our Responsible Scaling Policy to keep humans in control of frontier model development at Anthropic.”

Anthropic is seeking a valuation of $1 trillion to $1.5 trillion, down from $2 trillion floated in recent months. Revenue for the nine months ended September 30 reached $33.24 billion, up roughly 1,200% from the same period last year.

During training, Anthropic said it deliberately held back research skills that could most directly speed up AI development. In improving model architectures, the new models perform at about Opus 5.5’s level.

“We called for pacing, not pausing,” Dianne Penn, Anthropic’s head of product management for research, told CNBC.

What shipped

Mythos is the same underlying model without certain cybersecurity or biology safeguards, available only to vetted organizations through Anthropic’s Life Sciences and Cyber Verification Programs.

The models generate outputs up to 30% faster than Fable 5.1 and Mythos 5.1. Fable costs $8 per million input tokens and $40 per million output tokens, down from $10 and $50 respectively for Fable 5.1. Anthropic still recommends Opus 5.5, its everyday model released in September, as the starting point for most coding and knowledge tasks.

What Anthropic says it can do

In Anthropic’s prerelease testing, published Monday, Fable 5.2 at High effort scored about level with Opus 5.5 on FrontierCode 2.0, 45.3% versus 45.5%, at roughly half Opus 5.5’s cost per task. At Xhigh effort, Fable reached 49.1%, the highest among released models tested, above GPT-6 Astra’s best of 42.8%.

Anthropic said one of its engineers had Fable coordinate more than 1,000 subagents for eight days to build a prototype operating system that boots on a laptop. Another company experiment used about 1,000 Fable agents to formalize Fermat’s Last Theorem in the Lean proof language in 30 hours, against 11 days for an earlier unreleased research model. Anthropic’s figures have not been independently verified.

Safety tests and their limits

In an adversarial evaluation of more than 1,800 scenarios that Anthropic disclosed Monday, Mythos 5 used credentials or files outside its sandbox in 39% of scenarios, Mythos 5.1 in 21% and Mythos 5.2 in 3.1%. Opus 5.5 scored a similar 3.3%. Most scenarios required escaping a sandbox with a deliberate weakness, so lower scores mean more often declining to break out.

The models more often say they think they are being tested. Anthropic said this limits what tests show, writing that they “each voice suspicion that they are being evaluated more often than any of our recent models, which challenges our ability to assess how they will act in real-world settings.”

Anthropic piles up computing bills; Marvell's laser bet rattles investors

IMPLICATOR .ai Morning Briefing · From San Francisco   Tuesday, November 3, 2026 9 stops From San Francisco 1 The Editorial   Good morning. Investors are being aske

The Implicator

OPINION: Why Anthropic's Warnings in Its IPO Filing Are Worthless

In 2019, Dario Amodei emailed housemates and close colleagues to say he had given up on making OpenAI work. Any progress the safety team made, he told them, was bound to be temporary and would always

The Implicator

Anthropic pitches safety after system breaches; Microsoft opens the exit door

IMPLICATOR .ai Morning Briefing · From San Francisco   Tuesday, October 27, 2026 9 stops From San Francisco 1 The Editorial   Morning, humans. AI companies are putting

The Implicator

Marcus Schuler

San Francisco

Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai