Anthropic logs all Fable Studio conversations for 30 days, manual review mandatory

Anthropic updated its terms of service to mandate data collection for users of Fable Studio, its multimodal reasoning model. The new policy, labeled "limited data retention and review" for Mythos-class models, requires the company to store both user prompts and model outputs for a minimum of 30 days. During that window, Anthropic staff will manually review flagged conversations for policy violations.

The policy applies to all Fable users without exception. Anthropic states it will not train on retained data unless a user explicitly opts in, with one carve-out: conversations flagged as suspicious may be used for safety tuning regardless of user preference. The company has not defined the threshold for "suspicious" or disclosed what percentage of conversations trigger manual review.

Retention and deletion

The 30-day retention window covers the full conversation history—text prompts, image uploads, and the model's multimodal outputs. Anthropic's policy notes that data may be held longer if a conversation is under active investigation for terms-of-service violations. Users cannot delete their data during the retention period, and there is no API flag to skip logging.

Fable Studio remains the only production-grade multimodal reasoning model with sub-second latency for complex visual tasks, leaving practitioners with limited alternatives. Anthropic has not announced whether the retention policy will extend to future Claude releases or remain specific to the Mythos tier.