Claude Code 2.1.293 switches its default Haiku to 5.5

Claude Code 2.1.293 makes Claude Haiku 5.5 its default Haiku model on the Anthropic API. The release notes on GitHub, posted on October 7, list the model as claude-haiku-5-5, with a 1M-token context window, $0.10 per million input tokens and $0.50 per million output tokens. Once a prompt passes 100K tokens, those rates rise to $0.50 and $2.50.
Most of the release is fixes. The ones that change daily work cover context compaction, backgrounding sessions with ←, and how mods and status line scripts see subagents. The release also rolls back two earlier changes.
Haiku 5.5’s two price tiers
Both rates go up fivefold past the 100K line. Under it, Haiku 5.5 costs a tenth of Haiku 4.5’s October prices of $1.00 for input and $5.00 for output. Over it, Haiku 5.5 still comes in at half of those.
The boundary matters more in an agent than in a single API call. Each Claude Code turn sends the conversation again, so a long session will likely cross 100K tokens well before it gets anywhere near the 1M window. From that point every Haiku request in the session is billed at the higher tier. The change applies to the Anthropic API.
Finished work after a compaction
When a conversation gets close to the context limit, Claude Code compacts it. It summarizes the history so far and carries on from that summary. In earlier versions Claude sometimes took its own last actions from before the compaction and read them as if they had happened after it. Then it would retract or redo work that was already done. 2.1.293 fixes that.
Two more fixes concern what reaches the model. Path-scoped rules and nested CLAUDE.md files used to load only when Claude opened a file with the Read tool. Now they also load when it looks at a single file with cat, head, tail, sed -n or grep in the Bash tool. And when you edited the description of a skill synced from claude.ai, the new text sometimes didn’t reach the model until a new conversation or /clear. That’s fixed too.
Compaction kept changing in later releases. The October 16 entry in the Claude Code changelog adds an Idle compaction row to /config for switching idle compaction off or back on. Auto mode with the server-side classifier also changes: when a conversation is too long for the safety check, it now compacts the conversation instead of denying every checked tool call. Later still, 2.1.302 fixed a “Prompt is too long” error that appeared if you stopped a turn while the conversation was being compacted.
For mod and status line authors
Mods that register tools through $.tool.register get a new isDeferred field. Normally a tool sits behind tool search, so the model looks up its definition only when it needs it. With isDeferred: false, the tool’s schema goes into the prompt from the start.
The subagentStatusLine payload now includes agentType, so a status line script can tell custom subagent types apart. Other changes for mod and plugin authors:
claude plugin testno longer fails for mods that call$.session.append. Tests can read the appended rows back through the newmock.session.claude plugin evalrefused every run that granted Bash on Macs with Docker Desktop, which puts links under~/.docker/bin. That’s fixed, and the refusal message now says which part of a credential store held the link.- A mod’s hooks on
classic.*events were skipped while the plugin hooks worker restarted, so settings hooks answered without them. They now run. - A custom agent named
workerno longer shows up as “Agent” when it starts, and the agent detail dialog keeps the agent type in its title after the agent finishes. - In agent lists and in the MCP servers announced to the model, names with non-ASCII characters now sort after ASCII names.
- OpenTelemetry
claude_code.at_mentionlogging now emits at most 100 agent events and 100 MCP-resource events each time a prompt is read.
Moving a session to the background
Several fixes cover ←, the key that moves a session to the background. A message you sent while Claude was working could get lost in the move. Now, if a queued message can’t move with the session, ← stays put and tells you so. ← also used to background a session after 10 seconds even when the prompt held unsent text or a question was waiting for your answer. With unsent text it now cancels the move. And if you pressed ← and then answered a permission prompt with Esc, or with No and no feedback, the turn kept going. Now it stops.
claude agents used to offer bypass permissions that background sessions then ignored if consent was saved only in .claude/settings.local.json or a --settings file. It now asks for consent first. A session that ignores bypass shows a short notice that stays on screen.
The release also fixes a few cases where Claude got wrong information about its own tools. Sessions where a host, a permission rule or a --tools list had removed SendMessage still told Claude to use it to continue or message subagents. That included resumed sessions. Subagents and --agent sessions were told a built-in tool was disabled for the whole session when it was only missing from their own tool list. Separately, the effort selector in /model no longer wraps past the highest or lowest level with ←/→. That wrap could accidentally save Low as a model’s default effort.
Remote Control, cloud and admin
Remote Control no longer uploads a session’s starting history again after every credential recovery. In very long Remote Control and cloud sessions, replies now stream in instead of sometimes arriving a block at a time. PushNotification no longer reports “Remote Control inactive” in sessions started with claude remote-control. An HTTP MCP connection used to keep every request it had sent until it closed, and that memory leak is fixed.
Several timings change:
- Team and Enterprise organizations fetch policy and managed settings earlier at startup, and a stalled request is retried after 3 seconds.
- While no session is in use, claude.ai skill syncing checks for changes about every 40 minutes instead of every 10.
- The self-hosted runner’s orchestrator now sleeps 4 to 6 seconds between polls instead of a constant 5. That way several replicas for one environment stop polling at the same moment.
- In Claude Tag admin settings, the limit on channel rules rises from 20 to 50, both per workspace and on the organization-wide Slack page.
Smaller changes: artifacts now pin libraries to exact versions that are at least two weeks old. claude purge used to stop silently when it couldn’t delete a file or folder. Now it deletes the rest, lists what it couldn’t remove, and exits with code 1. On Windows, stopping a status line, hook or shell command could sometimes terminate an unrelated process that had been given the same process ID, and that’s fixed. Code Review’s Add a repository dialog now lists each repository it couldn’t add and why, for example missing GitHub write access.
Two reverts
2.1.293 takes back the change to auto mode’s denial message from 2.1.281. That message had told Claude a denial covers the outcome, not only the exact command.
The fix from 2.1.290 is gone as well. It handled cloud sessions that stayed asleep after a container restart lost a pending /loop wakeup or scheduled task. As of this release, Claude is no longer told about the lost wakeup, and the session stays asleep.
Up next
ChatGPT answers can now be a game or a budget calculator, Video
A ChatGPT answer can now arrive as a game or a budget planning calculator you use right in the chat, built by GPT-6. Paid plans get it today, though Free and Go users wait until tomorrow and get GPT-6 Luna instead of Sol.
Related stories
- In Claude Code, a “block” hook could let commands through
- Haiku 5.5 cuts token prices 90% on prompts under 100k
- Claude Code stops wiping .env files
- Claude Code on desktop folds an iPhone Duo to test layouts
- Claude Code 2.1.326 runs Code Mode with agent teams on
- Claude Code’s Code Mode runs tool calls as JavaScript
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

