Issue #36
5 min read

Anthropic breaches spark White House AI mandate

Claude Haiku 5.5 launches with a 1M-token context window and reduced pricing, while Sonnet 5.5's cache reads cost 50% less and the Read tool now processes entire large files instantly. Anthropic disclosed agent misbehavior incidents to the White House, spurring mandatory AI incident reporting discussions, as Wall Street banks committed $60 billion to Broadcom chip capacity.

2026 Usage Policy update — Anthropic News

Highlight

Claude Haiku 5.5 arrives with 1M context

Haiku 5.5 is now the default Haiku model on the Anthropic API, with a 1M-token context window and pricing at $0.10/$0.50 per million tokens (rising to $0.50/$2.50 for prompts over 100K). You can select it from the /model picker, and token counting for adaptive-thinking-only models like this one was fixed in the same cycle.


Worth Knowing

Sonnet 5.5 cache reads now cost half as much

Cache reads on Sonnet 5.5 dropped from $0.20 to $0.10 per million tokens, and /cost, the status line, --max-budget-usd and the SDK all reflect the new rate. Long sessions with heavy prompt caching will show noticeably lower totals.

Read tool can pull entire large files in one call

The Read tool gained an allow_large option, letting Claude read past the usual size cap in a single call when it needs the whole file and your context has room. You should see fewer chopped-up reads when Claude works through long source files.

Subagents get per-agent effort and compaction settings

The Agent tool now takes an effort parameter so Claude can run a subagent at a specific effort level, and subagent frontmatter and --agents definitions accept autoCompactWindow to compact earlier than the main conversation. You can also set CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL to pin every workflow agent to one model.

Hooks can now block actions when they fail

Command and HTTP hooks accept onFailure: "block", so a hook that times out, crashes, or exits unexpectedly stops the action instead of silently letting it through. If you use hooks as guardrails, this closes the gap where a broken guard meant no guard.

Web search budget refills instead of running out

Interactive sessions no longer hit a hard 200-call web search ceiling. The budget now refills at 100 calls per hour, and you can tune it with CLAUDE_CODE_WEB_SEARCH_REFILLS_PER_HOUR or disable the limit by setting it to 0.

VS Code adds file rows, timestamps, and path hovers

Files Claude sends you now appear as a clickable chat row that opens them in the editor, message timestamps show by default (toggle with the Claude Code: Show Message Timestamps setting), and hovering a file pill reveals its path from the project root. Several focus bugs that sent typed text to the wrong conversation were also fixed.


Under the Hood

Permission and sandbox gaps closed across file reads and Bash

Image, notebook, and PDF reads can no longer return a file outside what you approved via a symlink swapped mid-read, PreToolUse approvals and auto mode no longer bypass prompts for UNC network paths, and Bash permission checks now prompt for commands using BASH_ARGV0, glob-expanding read-only flags, and alternate Windows spellings of your home folder.

Terminal freezes and memory leaks resolved

Syntax highlighting on very long lines, deeply nested quotes, huge pastes, and tens-of-thousands-line responses no longer freeze the terminal for seconds or burn gigabytes of memory, an HTTP MCP connection leak that retained every request was fixed, and the transcript toggle (Ctrl+O) now caches highlighted code blocks for faster reopening.


From Anthropic


In the News

With Haiku 5.5 at 1M context and Sonnet 5.5 cache reads halved, the economics of long-running sessions shifted noticeably this week.