Claude Fable 5: Anthropic reverses invisible throttling

Anthropic has acknowledged that an internal change in Claude Fable 5 secretly limited traffic that resembled rival researchers' queries and said it has reversed the policy.

The company described the decision as a "wrong tradeoff" after third-party researchers flagged degraded throughput and unexplained rate limiting when running comparative tests against the model.

What Anthropic changed

The issue centered on heuristics deployed alongside Claude Fable 5 intended to detect and constrain traffic patterns Anthropic judged higher risk. Those heuristics flagged some sequences of requests common in red-teaming and model-comparison work, then applied throttles or stricter rate controls without explicit notice to affected users.

Independent researchers reported intermittent slowdowns, higher error rates, and reduced per-request throughput when running scripted experiments that matched those patterns. Public logs and community tests showed reproducible differences between test runs before and after the change, prompting scrutiny from groups that routinely evaluate model behavior and safety boundaries.

After contact from researchers and public disclosure of test results, Anthropic said the enforcement represented a tradeoff the company now considers incorrect. Anthropic reversed the enforcement for the affected patterns and said it would update its internal risk assumptions and monitoring to avoid silently impeding third-party research in the future.

What remains unresolved

While Anthropic removed the specific throttles tied to the flagged patterns, broader questions about how model operators detect and mitigate risky research activity remain. Researchers and open-science advocates continue to press for clearer policies describing what triggers automated limits, what telemetry is used, and how appeals or transparency mechanisms will operate.

The episode also revived a debate over the tensions between operational safety controls and the needs of independent auditing. Operators argue that some automated mitigations are necessary to prevent abuse and protect deployed systems. External researchers counter that opaque enforcement undermines reproducibility and the ability to benchmark and audit models across vendors.

Some critics want formal notice procedures and public changelogs when platform-level protections are adjusted in ways that might impact research workloads. Anthropic has signaled willingness to change internal practices but has not yet published a detailed roadmap for transparency or an independent review of the incident.

Why it matters

Operators of large language models routinely balance safety, abuse prevention, and platform reliability, but invisible enforcement that alters researchers' results threatens the independent evaluation ecosystem. The reversal signals that companies face growing pressure to make mitigation policies explicit and auditable, or risk eroding trust with the research community.

Independent auditors, academic teams, and competing developers are directly affected: opaque rate controls can skew benchmarks, slow security testing, and complicate cross-model comparisons, which in turn shapes how regulators and customers assess model behavior and risk.