Will Anthropic make a formal public ASL-4-equivalent capability-threshold determination before May 21, 2027? A YES requires a formal public statement that a model crosses any next-tier threshold (CB-2, Cyber Tier 2, or automated AI R&D), or activation of equivalent next-tier safeguards.

probability: 21%

rationale:

As of August 2026, no Anthropic model has triggered a next-tier (ASL-4-equivalent) capability threshold, and recent determinations have explicitly ruled models below these lines. The Claude 5 family—including Fable 5, Mythos 5, Sonnet 5, and Opus 5—was formally assessed as not crossing the CB-2, Cyber Tier 2, or automated AI R&D thresholds 55 sources. However, Mythos 5 was described as "near the border" for the novel chemical/biological (CB-2) threshold, representing a "much less clear and obvious judgment than with previous models" www-cdn.anthropic.com. Rather than declaring a threshold crossed, Anthropic responded by layering novel "ASL-3-plus" operational mitigations—such as fallback classifiers and 30-day retention—while keeping the models formally below the next-tier designation 2 sources.

The strongest headwind against an affirmative declaration is Anthropic's own security infrastructure timeline. The Frontier Safety Roadmap targets its flagship security goal—protecting model weights against high-end adversaries, which is a prerequisite for safely hosting next-tier models—for July 1, 2027 anthropic.com. Because this lands six weeks after the May 21, 2027 deadline, declaring a threshold crossing before then would effectively force Anthropic out of compliance with its own security mandates 2 sources. This structural reality incentivizes the company to lean heavily on interim operational safeguards or to refine threshold definitions. Evidence suggests Anthropic has already employed the latter strategy, repeatedly revising its Responsible Scaling Policy (RSP) thresholds between April and July 2026 to "better track the threat model of concern," a pattern that actively delays formal crossings 3 sources.

Government and regulatory dynamics introduce further disincentives for unilateral catastrophic-risk declarations. While Executive Order 14409 establishes a voluntary framework and disclaims mandatory preclearance whitehouse.gov, the regulatory environment heavily penalizes alarming public disclosures. When Anthropic released its June 2026 safety findings for Fable 5 and Mythos 5, a Commerce export-control directive forced a temporary worldwide shutdown of those models anthropic.com. Following their restoration on July 1, Anthropic's posture shifted toward co-developing consensus industry frameworks and voluntary standards rather than making unilateral risk determinations anthropic.com. This chilling effect significantly raises the political and commercial costs of issuing an ASL-4-equivalent statement.

There are, however, arguments in both directions. The rapid cadence of frontier model releases means 2 to 4 additional launches (such as a rumored "Mythos 6") are plausible before the May 2027 deadline. Given how close Mythos 5 already sits to the CB-2 line, a single capability jump could force an affirmative crossing determination www-cdn.anthropic.com. Anthropic's organizational culture also favors transparency and precaution; they previously activated ASL-3 safeguards for Opus 4 without a definitive threshold crossing. It remains possible that rapid capability progress—perhaps coupled with external pressure or another high-profile jailbreak incident—could force the company to formally activate next-tier safeguards on a precautionary basis over the next nine months.

Weighing these factors, a formal ASL-4-equivalent determination or the activation of equivalent next-tier safeguards before May 21, 2027, is a distinct possibility but ultimately unlikely. The combination of structural security timelines extending past the deadline, a demonstrated preference for layering ASL-3-plus mitigations rather than tripping thresholds, and severe regulatory disincentives strongly suggests Anthropic will maintain its current posture.