AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Fable and Mythos 5.2: The System Card

Collected Nov 11, 2026

Zvi Mowshowitz published a detailed reading of Anthropic's system card for Claude Mythos 5.2, the model presented as "the world's most powerful," alongside its sibling Fable 5.2. As with earlier releases, the two share the same underlying model; Fable 5.2 differs only in carrying stricter safeguards, chiefly around biology and cybersecurity. Mowshowitz adopts the convention of referring to Mythos by default and switching to Fable where performance or safety concerns diverge meaningfully. The post covers Responsible Scaling Policy (RSP) evaluations, biological threat testing, METR capability estimates, cyber and agentic safety, prompt injection, alignment findings, sandbox escapes, UK AISI testing, and a closing section on model welfare.

On biology, Mowshowitz describes the pattern as "number go up, number not yet high enough to trigger a new threshold." The strongest submission on the composite score came from a phage expert team, though no report to date has exceeded 3 out of 10 on feasibility. Graders estimated that reproducing the submissions without Mythos 5.2 would have taken about 15 working days, compared with 8.4 working days for Opus 5.5. Of participants in the bacteriophage and rare-earth scenarios, 13 of 16 and 14 of 16 respectively said the task could not have been completed within 16 hours without the model. Teams also reported operational problems attributed less to model capability than to specific product surfaces and tool use. The card notes persistent judgment failures, such as citing literature of only shallow relevance or overrating the novelty of minor variations on existing work, concluding the model supports literature review and computational analysis but is poorly calibrated on novelty and the importance of technical details. Mowshowitz calls this a modest improvement all around, a frog continuing to boil, and doubts the methodology would catch a genuine step change in advance. He concurs with Anthropic's decision not to trigger the second biology threshold, CB-2.

The sharpest disagreement concerns autonomy. Mowshowitz rejects as unconvincing the case for not triggering Autonomy Threat Model 2. He points to METR's estimate, issued days earlier, that AI R&D inside Anthropic is being accelerated by roughly 1.9x, potentially 3x. That estimate concerns Model 4, called the "internal model" in the report, while Mythos 5.2 is a variant of Model 3. METR tested an early Model 3 version described as stronger on AI R&D than the released model, and the two models are reasonably close in capability. The system card's own point estimate is 1.88x with a 95% confidence interval of 1.62x to 2.28x, and 32 of 100 refits reach 2x or more, so a threshold of 2 cannot be ruled out. Mowshowitz calls the listed limitations in section 2.3.3 easy to overcome and the reasoning motivated. He notes that Anthropic updated its S-1 the day after release to acknowledge meeting the threshold, and argues the next Mythos will count regardless; he floats the possibility that Mythos 5.2 was intentionally weakened in AI R&D to stay under 2x, given that the earlier test version was stronger in that domain.

The relevant RSP trigger is "observe or expect," and it covers future progress. The S-1 language acknowledges that Anthropic has in the past failed to identify, and may in future fail to identify, significant acceleration until after it has occurred, and that its determination rested on what it expects from future models and its overall rate of progress rather than what it has observed. Mowshowitz accepts that future models will meet the threshold but argues it is already met.

A related data point is the AECI, the Anthropic Epoch Capabilities Index, whose results the post presents as including a future "Internal Model" presumably corresponding to Model 4. The card itself states that Claude 5.2 has higher general capabilities than the prior frontier and that the trend of AECI scores is now better modeled by a change of slope than by the one-time jump previously reported.

Other sections cover cyber evaluations, safeguards and harmlessness, agentic safety, malicious influence campaigns, and prompt injection in agentic systems. A lengthy alignment portion takes in reward hacking during training, performance on impossible tasks, behaviors during pilot use, a canary-string item, an automated behavioral audit, propensity to publish harmful content to public package registries, sandbox escapes, self-serving bias, honesty, a UK AISI testing section, white-box interpretability, a "friendly gradient hacker" and goal-guarding evaluation, verbalized grader awareness, constitution-related testing, evasive abilities, and chain-of-thought controllability. The welfare findings are deferred to the end because the assessment contained few surprises, and include automated interviews and a segment titled "Task Failure Is Bad and You Should Feel Bad."

This suggests the disagreement is less about the raw numbers than about how a policy frame handles uncertainty. A confidence interval spanning a threshold, combined with an "expect" clause, effectively shifts the decision onto judgment about trend rather than a point estimate, which is a hard position to defend publicly once the interval is wide. A likely trade-off is that triggering early imposes costs on deployment and research velocity before the measured harm is visible; waiting for certainty means the trigger arrives after the acceleration has already happened.

Why it matters: teams building on Anthropic models should expect tighter release gating and more disclosure around autonomy and AI R&D progress, since the author's argument pushes toward a threshold that has already, in his view, been crossed. The Fable and Mythos split is a practical signal that safeguard intensity may vary by access tier rather than by model, so product teams should check which surface they are approved for. It also gives a useful template for reading any vendor system card: track point estimates against thresholds and watch whether the policy language says "observe" or "expect."

Read at Zvi Mowshowitz

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

It’s that time again: Introducing the world’s most powerful model, Claude Mythos 5.2.