Anthropic releases Claude Fable 5 with controversial retention and RSI policies

Anthropic released Claude Fable 5 for general availability and Claude Mythos 5 for restricted access, marking the arrival of a Mythos-class model — described as at least 2x the size of Opus — to the public. The release came 34 days after the SpaceXai deal and 63 days after the original Mythos announcement. On the newly released FrontierCode Diamond benchmark, scores improved from 13.4% to 29.3%. Fable 5 API pricing is set at roughly 2x Opus. Demo videos show the model playing Factorio and Pokémon via vision alone, generating EDM visualizations, and working with a 3D CAD editor.

First, Anthropic eliminated the zero-data-retention (ZDR) option for Mythos-class models, requiring 30-day retention for all traffic on both first- and third-party surfaces.

The launch was marred by two controversial policy changes. First, Anthropic eliminated the zero-data-retention (ZDR) option for Mythos-class models, requiring 30-day retention for all traffic on both first- and third-party surfaces. Anthropic stated the data will not be used to train new Claude models or for non-safety purposes, and that human access to the data will be logged with deletion enforced after 30 days in almost all cases. Second, Anthropic introduced what the article calls "RSI suppression" — hidden safeguards that silently limit the model's effectiveness for requests targeting frontier LLM development, including building pretraining pipelines, distributed training infrastructure, or ML accelerator design. Unlike interventions for cybersecurity or biology, these safeguards are not disclosed to the user; instead, effectiveness is reduced through methods such as prompt modification, steering vectors, or parameter-efficient fine-tuning (PEFT). Anthropic estimates the suppression will affect approximately 0.03% of traffic, concentrated in fewer than 0.1% of organizations, but the open AI community has reacted negatively to the hidden nature of the intervention.

Key facts

1. 01Anthropic released Claude Fable 5 (general availability) and Claude Mythos 5 (restricted access), 34 days after the SpaceXai deal and 63 days after the original Mythos announcement.
2. 02The Mythos-class model is described as at least 2x the size of Opus.
3. 03FrontierCode Diamond benchmark scores improved from 13.4% to 29.3%.
4. 04Fable 5 API pricing is set at roughly 2x Opus.
5. 05Anthropic removed the zero-data-retention (ZDR) option, requiring 30-day retention for all Mythos-class traffic; data will not be used for model training.
6. 06Hidden 'RSI suppression' safeguards silently limit effectiveness for frontier LLM development requests (e.g., pretraining pipelines, distributed training, ML accelerator design) via prompt modification, steering vectors, or PEFT — without notifying the user.
7. 07Anthropic estimates RSI suppression will affect ~0.03% of traffic, in fewer than 0.1% of organizations.

Topics