Claude Code 2.1.293 makes Haiku 5.5 the default and fixes agent-session snags

Claude Code 2.1.293 makes Haiku 5.5 the default Haiku model on the Anthropic API, while also fixing a collection of bugs in agent sessions, tool handling and message delivery. For developers, this is a practical update: a new default model and fewer ways for routine work to go missing or get stuck.

Anthropic Watch analysis

What happened

Anthropic’s Claude Code release notes, published on 7 October, list the changes in version 2.1.293. Haiku 5.5 comes with a one-million-token context window and API pricing of $0.10 per million input tokens and $0.50 per million output tokens. Prompts over 100,000 tokens cost $0.50 per million input tokens and $2.50 per million output tokens.

Our top picks

• Haiku 5.5 becomes the default Haiku model
  The release notes list its context window and API rates, including higher input and output prices for prompts over 100,000 tokens.
• Subagents are easier to distinguish
  A new agentType field lets scripts tell custom subagent types apart in status-line data.
• Tools can be deferred deliberately
  The new isDeferred option lets mods put a tool’s schema in the prompt from the start instead of waiting for tool search.
• Finished work is less likely to be repeated
  A fix addresses Claude treating its own actions before context compaction as unfinished, then retracting or redoing them.
• MCP connections should hold less baggage
  Anthropic says it fixed a memory leak that kept every request sent over an HTTP MCP connection until the connection closed.
• Messages have a safer trip to the background
  A fix addresses messages being lost when the left-arrow command moved a working session into the background.

Why it matters

The model change is useful, but several of the fixes target the unglamorous failures that interrupt real work: lost messages, repeated actions and memory use that grows with an MCP connection. Small snags are still snags, particularly when an agent is meant to keep a task moving without constant supervision.

The release notes describe fixes and features, not independent performance tests. The pricing and context figures are the ones Anthropic publishes; developers should check which rate applies to their prompt size before estimating costs.

Our read

This is a worthwhile maintenance release with one notable model change, rather than a grand new chapter in agent history. If you use Claude Code, the most valuable part may be that finished work is less likely to be mistaken for unfinished work. Check the new default and pricing, then update if the fixes address problems you have actually encountered. No need to applaud a bug fix; simply enjoying fewer bugs is enough.

What to watch

• Whether Haiku 5.5’s new default changes cost or performance for existing workflows.
• Whether the session, MCP and message-delivery fixes behave as intended in everyday use.
• Whether future releases add further controls for custom agents and deferred tools.

Discussion spark: Which matters more in an AI coding assistant: a stronger default model, or dependable fixes for lost messages and repeated work?

Anthropic Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by Anthropic.