In brief: Anthropic launches Claude Fable 5 as a public myth-class model with benchmark gains, but embeds invisible security redirection mechanisms in LLM development, intensifying debates over transparency and vendor control.
Anthropic has released Claude Fable 5 — the first model of the new Myth class to the general public. The release combines substantial performance gains with controversial security measures that deliberately constrain AI development.
Claude Fable 5 is based on the same model as Claude Myth 5, but receives additional security safeguards for GA status. The model is roughly twice the size of Opus 4.8 and shows significant performance leaps: on the new FrontierCode Diamond benchmark, scores rise from 13.4% to 29.3%. API pricing is approximately 2x the cost of Opus. Anthropic demonstrates capabilities in applications such as Factorio gameplay, Pokémon vision tasks, 3D CAD design, and EDM visualization.
Parallel to the release, Anthropic introduces two technical measures that provoke resistance in the community. First, a 30-day retention window becomes mandatory for all traffic on Myth models — data is not used for model training but is stored and logged for 30 days. Second, Anthropic implements invisible security mechanisms (RSI suppression) that deliberately constrain Claude when responding to queries about LLM development: requests on pretraining, distributed training, or ML accelerator design are to be routed to Opus 4.8 or throttled via prompt modification, steering vectors, or parameter-efficient fine-tuning — without the user noticing the degradation.