Claude Code sets Fable 5.1 as default model and adds a fullscreen diff panel

Claude Code shipped a dense cluster of releases in the first week of September 2026, from v2.1.251 through v2.1.260. The headlining change is that Claude Fable 5.1 is now the default Fable model in Claude Code. But there is plenty more worth knowing about, from a genuinely useful new panel for reviewing code changes to a bug fix that was silently inflating costs for anyone using Fable 5.1 heavily.

Fable 5.1 is now the default

Claude Fable 5.1 (claude-fable-5-1) replaces Fable 5 as the default model in Claude Code from v2.1.257 onwards. It brings a one-million-token context window, priced at $10 per million input tokens and $50 per million output tokens. Cache reads cost $0.25 per million tokens, which is one quarter of the input price and makes prompt caching significantly more attractive for long-running sessions.

Fable 5.1 is designed for long-horizon agentic work: extended coding tasks, multistep research, and working across documents, spreadsheets, and slides. If you were already using Fable 5, the upgrade is automatic. If you use the /model picker in Claude Code, the model is selectable as claude-fable-5-1. A bug that prevented the picker from showing Fable 5.1 for eligible organisations has been fixed; previously it only worked if you typed the model ID directly.

For GitHub Copilot users, Fable 5.1 is available on Pro+, Max, Business, and Enterprise plans from 1 September 2026. One detail worth noting: unlike other Claude models in Copilot, Fable 5.1 requires data retention by default so that Anthropic’s safety classifiers can operate.

A live diff panel for code review

The most visible new feature in v2.1.260 is a fullscreen /diff panel. When you are in fullscreen mode, typing /diff opens a panel beside the conversation that shows your repository’s uncommitted changes in real time as Claude edits files.

Before this, if you wanted to see what Claude was actually doing to your code while a task was running, you needed a second terminal open with git diff, or you had to flip back to your IDE. The new panel keeps that view in the same window without interrupting the conversation. It is a small change in concept but a meaningful one in practice for anyone doing extended coding sessions inside Claude Code.

Two related fullscreen improvements shipped alongside it: Ctrl+L or Cmd+K now clears the transcript view the way a normal terminal clear would, and the 1M-context auto-compact threshold has been adjusted so it kicks in shortly before the limit rather than at it.

The Fable 5.1 cache bug you probably did not know you had

This one matters more than it sounds. Before the fix in v2.1.260, prompt caching on Fable 5.1 was not covering the context attached after tool results. That meant every tool-call turn was resending the full context as uncached input, which at $10/MTok input adds up quickly during any task with frequent tool use.

If you ran heavy Fable 5.1 sessions between the model’s release and the fix landing on 1 September 2026, your costs for those sessions will have been higher than they should have been. The fix is now in place, and the new prompt-cache miss diagnostics (visible in /cost and the prompt_cache field in the status line) will tell you if something is still causing misses and give you a likely reason, such as a changed tool definition or a context that went idle past the TTL.

Headless session commands now discoverable

If you run Claude Code in headless or CI environments, the commands for managing background sessions (attach, logs, stop, respawn, and rm) are now listed in claude --help. Previously they existed but were not surfaced there, which made them easy to miss if you had not read the documentation carefully.

Alongside that, when a background session is already running and you try to resume it, Claude Code now tells you the exact command to use: claude attach <id> with the actual session ID filled in, rather than a generic message.

Remote Control and MCP fixes

Several bugs affecting Remote Control sessions have been resolved. When a phone, browser, or the claude.ai app attaches to a terminal session, the permission mode shown is now accurate rather than stale. Sessions that were stopped from a connected phone or browser were previously showing a stuck spinner and an active Stop button even after the turn had ended; that is fixed.

On the MCP side, /mcp reconnect and enable were incorrectly reconnecting servers that a managed MCP allow/deny list should have blocked. That has been corrected. A separate fix handles SDK and cloud sessions that were hanging indefinitely when a handshake acknowledgment was lost; the wait now times out after 70 seconds and marks only the affected server as failed rather than stalling the whole session.

A stricter rule for autonomous mode

Auto mode now includes a Containment Escape rule. Cloud metadata credential fetches, egress evasion, and cross-tenant access attempts are no longer auto-approved unless the environment explicitly flags them as expected. For most individual developers this will not change anything day to day, but for teams running Claude Code in shared or multi-tenant infrastructure it closes a meaningful gap.

What this means for you

For most Claude Code users, the practical effect of this release cluster is straightforward. You get a better model by default, a useful new panel for watching your code change in real time, and a fix for a caching bug that was costing you money you did not realise you were spending. Check your /cost output after a tool-heavy session if you want to confirm caching is working as expected.

For teams and CI operators, the combination of cleaner --help discoverability, the CLAUDE_CODE_SUBAGENT_MODEL_FORCE environment variable for locking subagent models, and the stricter auto-mode containment rules adds up to meaningfully better control over how Claude Code behaves in automated pipelines and shared environments.

The full changelog is available at code.claude.com/docs/en/changelog.