Claude Haiku 5.5
Claude Haiku 5.5 is the new default Haiku model
On the Anthropic API
- Context
- 1M tokens
- Per million tokens
- $0.10 in, $0.50 out
- Prompts over 100K
- $0.50 in, $2.50 out
Claude Haiku 5.5 is now Claude Code's default Haiku model on the Anthropic API. It takes a million tokens of context, at 10 cents in and 50 cents out per million, five times that above 100,000.
Work survives a compaction
>Claude no longer redoes finished work after a compaction
- Before
- Claude could retract or redo work it had already finished
- Now
- its last actions before a compaction stay done after it
After a context compaction, Claude sometimes misjudged its own last actions from before it. It could retract or redo work it had already finished. That's fixed, so long sessions pick up where they were.
Rules load on Bash file views
>Path-scoped rules load when Claude views a file through Bash
- Also loads
- path-scoped rules and nested CLAUDE.md files
- Commands
- cat, head, tail, sed -n or grep on a single file
Path-scoped rules and nested CLAUDE.md files now load when Claude views a file in Bash instead of the Read tool. That covers cat, head, tail, sed or grep on a single file.
Backgrounding with ←
>Sending a session to the background with ← keeps your message
- Queued message
- kept; if it can't move, ← stays put and says so
- Unsent prompt text
- cancels the move to the background
- Esc on a permission prompt
- still stops the turn right after ←
Sending a session to the background with the left arrow is safer. A message queued while Claude works is kept, unsent text in the prompt cancels the move, and Escape at a permission prompt still stops the turn.
Status lines and mods
>Status line scripts can tell custom subagent types apart
- subagentStatusLine
- the payload adds agentType
- Mods
- isDeferred: false lists a tool's schema in the prompt from the start
The subagent status line payload now includes the agent type, so your script can tell custom subagents apart. Mods can also list a tool in the prompt from the start, instead of behind tool search.
Rolled back
>Cloud sessions that lose a wakeup stay asleep again
- Cloud sessions
- a restart that loses a /loop wakeup or scheduled task leaves the session asleep
- Auto mode denials
- the message is back to its earlier wording
An earlier fix for cloud sessions is rolled back. If a container restart loses a pending loop wakeup or scheduled task, the session stays asleep and Claude isn't told. Auto mode's denial message reverts too.
Smaller fixes
>HTTP MCP connections stop leaking memory
- HTTP MCP
- a connection no longer keeps every request until it closes
- /model effort
- ← and → stop at the ends instead of wrapping to Low
- claude logs, stop, kill
- no longer sign you out when your login is expiring
HTTP MCP connections stop holding every request they've sent until they close. The model picker's effort arrows stop at the ends instead of wrapping to Low, and claude logs and claude stop no longer sign you out.
The full changelog
Every change is in the Claude Code changelog
CHANGELOG.md in anthropics/claude-code
The full list, including fixes for Claude in Slack and Code Review, is in the Claude Code changelog on GitHub.
