Anthropic explains Claude Fable 5's safety features amid criticism

Anthropic published a help article on June 10th (local time) explaining the mechanisms and behavior of its top-tier AI model "Claude Fable 5," which was made available to the public the previous day. The article details its protective functions (guardrails), how it handles blocked requests, and the management of its automatic switching feature.

Fable 5 is equipped with a mechanism that switches to the lower-tier model "Claude Opus 4.8" when it detects questions related to cybersecurity or biology. However, criticism has spread on X (formerly Twitter) that the model blocks even harmless requests.

According to Anthropic's explanation, Fable 5 performs automatic safety checks on all user requests and blocks them in three areas: "offensive cybersecurity techniques" such as the creation of exploits, malware, or attack tools; "biology and life science questions" including experimental methods and molecular mechanisms; and requests related to "model distillation."

These protective functions are intentionally set broadly and may inadvertently catch legitimate tasks such as authorized security testing or harmless biological research. During the announcement on the 9th, Anthropic acknowledged that blocking harmless requests "is a source of dissatisfaction for some users" and stated that they aim to reduce false positives by updating and improving safety measures after release.

According to the support page, the behavior after a block occurs varies depending on the setting of the automatic switching function. This function is enabled by default in Claude, Claude Cowork, Claude Code, Claude Design, and Claude for Microsoft 365. In this case, blocked requests are re-executed by "Opus 4.8" within the same conversation, and a notification is displayed indicating that the model has switched. After switching, the model selection remains Opus, but users can switch back to Fable 5 at any time. However, if they switch back without modifying the request, the original request may remain in the conversation, potentially leading to it being blocked again by the same protective function. Anthropic recommends editing the message before retrying.

The automatic switching function is enabled by default when Fable 5 is selected for the first time. It can be turned off by going to [Settings] -> [Features] (or [Config] -> [MODEL & OUTPUT] in Claude Code) and disabling "Switch model when message is flagged." When disabled, instead of switching models upon a block, the conversation pauses, allowing the user to choose between editing the message to retry with Fable 5 or manually sending it to Opus.

Billing was also explained. If a request is blocked before Fable 5 outputs a response, it immediately switches to Opus, and only Opus's rates will be charged. If the block occurs mid-response, the portion generated before the block is charged at Fable 5's rate, and the remainder is charged at Opus's rate.

For the Claude API, the behavior differs: automatic switching is not enabled by default, and API users need to configure the switching themselves.

Anthropic urges users to report instances where blocked requests appear unrelated to security or biology, or when legitimate work is repeatedly blocked, via the "Send Feedback" option. Reporting incorrectly blocked requests is said to help improve the accuracy of the protective functions. The company plans to consider opening up access for dual-use (military and civilian) cyber defense and biology research in the future, aiming to support legitimate biological and defensive cybersecurity work as its safety systems mature.