Cheaper models and longer contexts only help if you can control what runs where and what happens when a safeguard breaks. This week touches both sides: Haiku 5.5 becomes the default small model, sub-agents get their own effort and compaction settings, and hooks can finally fail closed instead of silently letting an action through.
Week 41 (October 5-9, 2026) covers versions 2.1.290 to 2.1.296: 7 releases and 564 changelog entries. Below are the ones that change how you configure models, agents, hooks and MCP connections.
What Haiku 5.5 changes in your model setup
A new default Haiku. Claude Haiku 5.5 (claude-haiku-5-5) is now the default Haiku model on the Anthropic API. The change landed in 2.1.293, and the details live in the model configuration docs.
What it offers. The context window is 1M. Pricing is $0.10/$0.50 per Mtok, and prompts over 100K cost $0.50/$2.50 per Mtok. That is a fivefold step up once a prompt crosses 100K, so long contexts on this model are priced differently from short ones.
Conditions.
- The changelog entry names the Anthropic API as the place where Haiku 5.5 is the default.
- The input says nothing about the default on other providers, so check your own setup before assuming it changed.
- If you pin a Haiku model explicitly, the entry does not say what happens to your pin. Look at your configuration before you rely on either behavior.
Why it matters in practice. Whatever you route to Haiku (cheap, fast sub-agent work, for example) now gets the 1M window. The price jump above 100K makes the sub-agent compaction setting in the next section worth a look. I can’t tell you the right window value for your workload, because the changelog doesn’t give one.
Changes for everyday work
Sub-agents get their own effort. The Agent tool now takes an effort parameter, so Claude runs a sub-agent at the effort level you ask for (2.1.292). You can say in the prompt that a cheap lookup task should run at a lower level than the main task. See the sub-agents docs.
Sub-agents can compact earlier. Sub-agent frontmatter and --agents definitions accept autoCompactWindow (2.1.296). A sub-agent can then auto-compact earlier than the main conversation’s window. Combined with Haiku 5.5’s pricing tiers, this is the setting to test if you want sub-agent prompts to stay small.
Hooks can fail closed. Command and HTTP hooks accept onFailure set to block (2.1.295). A hook that can’t start, times out, or exits with an unexpected code then blocks the action instead of letting it through. For a hook that enforces policy, this removes a quiet gap: a broken script no longer means an unchecked action. Details are in the hooks docs.
Fast mode in Slack. Mention Claude with !fast to switch a Slack thread to fast mode (2.1.290). The thread moves to Opus if needed. !fast off switches back, and replies show (fast) while the mode is on. The fast mode docs cover the rest.
Technical notes
Terminal. Claude Code supports the Program Status Protocol (OSC 7501) as of 2.1.295. Terminals that implement it can show whether Claude Code is working, waiting on you, or done. Whether your terminal does is something to check in its own documentation; see also the terminal config docs. The /copy picker now includes quoted text, so a drafted message copies without its > markers. claude attach <name> and claude logs <name> accept part of a session name instead of the id.
MCP protocol. Local (stdio) MCP server connections negotiate protocol version 2026-07-28 by default on every install, including Bedrock, Vertex and Foundry (2.1.292). claude.ai connectors do the same on installs that fetch no flags (2.1.295). MCP_PROTOCOL_NEGOTIATION=legacy opts out in both cases. Stdio servers that ignore the newer check are slow to connect once, then remembered for 7 days and connected the older way without the wait. The MCP docs have the background.
MCP limits. The default limit on MCP tool descriptions sent up front and on MCP server instructions rose from 2,048 to 4,096 characters. Descriptions loaded through tool search are now cut at 16,384.
Tools. The Read tool has an allow_large option to read a text file past the usual size limits in one call. WebFetch now says how much page text was unread past 100,000 characters and takes an offset to read on. The interactive WebSearch budget refills at 100 calls/hour instead of ending after 200 calls; CLAUDE_CODE_WEB_SEARCH_REFILLS_PER_HOUR sets the rate, and 0 turns the refill off.
Agents and retries. CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL runs every workflow agent on one model while other sub-agents keep theirs. CLAUDE_CODE_OVERLOADED_RETRY_BASE_DELAY_MS and CLAUDE_CODE_OVERLOADED_RETRY_MAX_DELAY_MS set the base and maximum backoff delay when retrying overloaded (529) requests.
Plugins. claude plugin install --marketplace <source> adds the marketplace if needed, under the same policy checks as marketplace add, then installs the plugin from it.
VS Code.
- Message timestamps show by default; the Claude Code: Show Message Timestamps setting turns them off.
- Files Claude sends you get a chat row. Click the file name to open it in the editor, with Claude’s caption underneath. The row stays visible in Focus view.
- Claude in Chrome now asks before browser actions in every session, including one connected with
@browser. Allowing a site for the session stops repeat asks.
What to take away
If you work in the terminal.
- Check whether your terminal implements OSC 7501 and shows Claude Code’s working, waiting and done states.
- Try the
/copypicker with quoted text, and theallow_largeoption on Read for oversized text files. - If you use WebSearch heavily, look at
CLAUDE_CODE_WEB_SEARCH_REFILLS_PER_HOURnow that the budget refills hourly.
If you run parallel sessions and sub-agents.
- Try
claude attachandclaude logswith a partial session name. - Set
efforton sub-agent calls where a task doesn’t need the main level. - Test
autoCompactWindowon a sub-agent that runs on Haiku 5.5, especially if its prompts approach 100K. - Use
CLAUDE_CODE_WORKFLOW_SUBAGENT_MODELif you want all workflow agents on one model.
If you administer integrations and company deployments.
- Review your policy hooks and decide which should use
onFailureset toblock. Test what happens when the hook script is missing or times out. - Check whether any local stdio MCP server or claude.ai connector misbehaves under protocol version 2026-07-28.
MCP_PROTOCOL_NEGOTIATION=legacyis the opt-out. - Confirm which Haiku your configuration resolves to, since the new default is documented for the Anthropic API.
- Note that
claude plugin install --marketplaceapplies the same policy checks asmarketplace add, and that Claude in Chrome in VS Code now asks before browser actions in every session.
Sources: the Claude Code changelog, versions 2.1.290 to 2.1.296 (October 5-9, 2026), and the linked Claude Code documentation pages. Availability of features tied to a specific surface, such as Slack or VS Code, may change or differ by setup.