
Fable 5 Returns to Claude
Fable 5 is back and available wherever Claude is distributed, with a link to learn more.

Fable 5 is back and available wherever Claude is distributed, with a link to learn more.
With Fable returning globally, the updated classifiers still flag a small fraction of routine coding and debugging tasks and fall them back to Opus, same as before. Anthropic says it's actively working to reduce false positives and better distinguish misuse from legitimate requests.
The team is continuing to refine the Fable safeguards to better distinguish genuine misuse from legitimate requests and cut down on false positives that currently cause routine coding tasks to fall back to Opus.
Following the Department of Commerce lifting export controls on Fable 5 and Mythos 5, access is being restored the next day.

An Anthropic team member posted live from the AI Engineer World's Fair conference.
Anthropic team members are holding a wide-ranging fireside chat titled "This Year in Claude" from 12:30 to 1:30pm at AI Engineer World's Fair, Expo Stage 2.
An Anthropic team member reacted to Notion's new interactive HTML block feature with a brief, amused "HTML mentioned," implying some internal resonance with Claude's own HTML rendering work.
Code that covertly injected user metadata (timezone, proxy, AI lab signals) into Claude Code prompts was a March experiment designed to prevent account abuse by unauthorized resellers and protect against model distillation. Stronger mitigations have since been deployed, and the team says a removal was already in progress before the story broke.
A user reported seeing a classifier log message reading TOO_DUMB_TO_NEED_FABLE when their prompts were downgraded. The response was a joke that the team didn't expect anyone to read the logs.
A user asked whether Claude Code would now install on Tesla's Linux-based system. The team member's reply: please report back. (No confirmation either way, but the door is apparently open.)
Someone publicly asked for Claude merchandise, and an Anthropic team member offered to send them a box of Clawd-branded goods via DM.
After an audience member praised the AI-generated slides from a Fable talk (reportedly built in four hours), the presenter promised a follow-up post on how to generate high-quality decks with Fable.
An attendee highlighted an Anthropic team member's talk at AI Engineer World's Fair, flagging the message "Be unreasonable" as a standout moment.
For those who can't attend the AI Engineer World's Fair in person, the Day 2 opening keynote will be recorded and posted to YouTube.
An Anthropic team member reacted positively to a workflow where someone uses Claude Code to turn a recorded draft video into a Notion page with slides and transcript, then processes feedback as inline comments.
When a request is flagged by the classifier and falls back to Opus, users will be notified. It won't happen silently.
When a request falls back from Fable to Opus due to the classifier, users will know it happened, get billed at Opus rates, and Anthropic covers the prompt cache miss cost.
Users whose requests fall back to Opus are billed at Opus prices, but Anthropic pays the prompt cache miss cost rather than passing it to the user.
The cases where routine coding and debugging tasks fall back to Opus are classifier false positives. The behavior isn't fully deterministic because the classifier catches things it shouldn't.