Claude Fable 5.1 now default model

Highlight

Claude Fable 5.1 becomes the default Fable model

Fable 5.1 ships with a 1M-token context window at $10/$50 per million tokens and $0.25 per million for cache reads, and it is now what fable resolves to. Auto-compaction was tuned so 1M-context sessions compact shortly before the limit, prompt caching now covers context attached after tool results, and changing /effort mid-session no longer invalidates the cache. If you run through a Claude apps gateway, fable and best still resolve to Fable 5 until gateways are configured for it -- pick Fable 5.1 explicitly in /model.

Worth Knowing

Live diff panel opens beside the conversation

In fullscreen mode you can toggle /diff to open a panel showing your uncommitted changes as Claude edits them, so you can watch a refactor take shape without switching to your editor or running git diff yourself.

Prompt word-editing keys now match Bash

Ctrl+W deletes back to whitespace, Alt+F and Alt+D stop at word end, and punctuation counts as a word separator, matching Bash readline behavior. The keybindingFlavor setting no longer does anything, so if you had it configured you will notice the change immediately.

New skill-doctor command shows unused skill overhead

/skill-doctor lists which loaded skills go unused and how much context each one costs, giving you a concrete basis for pruning. Related diagnostics also improved: /cost and the status line's prompt_cache field now name a likely cause when a cache miss happens.

Headless sessions gain no-prompt mode and larger output limits

--permission-prompts none automatically denies anything that would prompt while your active permission mode keeps deciding, which makes unattended hosts predictable. New bashOutputMaxChars and taskOutputMaxChars settings raise how much command and background-task output reaches Claude inline, up to 128K characters. If you are building scheduled automation on top of this, the scheduled repository reviewer cookbook walks through a read-only agent that resumes its own session.

VS Code adds archiving, a model pill, and MCP management

The session list now archives instead of deletes, with an Archived sessions group and Unarchive action, plus Active and status filters for finding work in progress. The input footer gained a model pill showing your current model and effort level, and the MCP servers dialog can now add and remove servers without leaving the IDE.

Auto mode asks before reading outside your working directories

The first time auto mode reads a file outside your working directories, you get a one-time prompt with the option to block such reads via permissions.blockReadsOutsideWorkingDirectories. A new Containment Escape rule also stops auto mode from approving cloud metadata-credential fetches, egress evasion, and cross-tenant access.

Under the Hood

Permission rules with parentheses and bad patterns no longer misfire

File rules whose paths contain parentheses were being dropped as invalid, leaving supposedly read-only folders writable, and a single uncompilable pattern could make every edit fail with Invalid regular expression. Bash checks also no longer auto-approve command substitutions hidden in zsh variable assignments or certain [[ ]] conditionals.

Faster streaming, typing, and terminal resize

Long conversations do less re-render work per turn, streaming no longer slows as a reply grows, already-rendered blocks are skipped by layout, and typed characters no longer land out of order during fast input; idle CPU in -p and SDK sessions also dropped.

From Anthropic

In the News