Haiku 5.5 Arrives with a One Million Token Context Window

I'm Shannon, and this is the Claude Notes Brief -- Claude Code updates and Anthropic news for the week of Monday, October twelfth. Claude Haiku five point five brings a one million token context window. Sonnet cache reads cost half as much. And the White House weighs mandatory AI incident reporting.

Let's start with Claude Code, where Claude Haiku five point five arrived as the new default Haiku model on the Anthropic API, carrying a one million token context window. Pricing sits at ten cents per million input tokens and fifty cents per million output, rising to fifty cents and two dollars fifty for prompts over one hundred thousand tokens. You can pick it from the model selector in the CLI, and token counting for adaptive-thinking-only models like this one was corrected in the same cycle. The cost story continues on Sonnet five point five, where cache reads dropped from twenty cents to ten cents per million tokens.

The cost command, the status line, the budget flag, and the SDK all reflect the new rate, so long sessions with heavy prompt caching will show noticeably lower totals. With that much context available, the Read tool got a matching change -- it can now pull an entire large file in a single call when Claude needs the whole thing and your context has room. Expect fewer chopped-up reads when Claude works through long source files. Subagents picked up finer controls too.

The Agent tool now accepts an effort level, so Claude can run a subagent at a specific setting, and subagent definitions can compact their context earlier than the main conversation. There's also an environment variable to pin every workflow agent to a single model. On the guardrails side, command and HTTP hooks can now block an action when the hook itself fails. If a hook times out, crashes, or exits unexpectedly, the action stops instead of quietly going through -- closing the gap where a broken guard meant no guard at all.

Web search got more forgiving as well: interactive sessions no longer hit a hard two hundred call ceiling. The budget refills at one hundred calls per hour, and you can tune that rate or turn the limit off entirely. And in Visual Studio Code, files Claude sends you now show up as a clickable chat row that opens them in the editor, message timestamps display by default, and hovering a file pill reveals its path from the project root.

Under the hood, this was a week about permissions and performance. On permissions, image, notebook, and PDF reads can no longer return a file outside what you approved by way of a symlink swapped mid-read. Approvals and auto mode no longer skip prompts for UNC network paths, and Bash permission checks now prompt correctly for a handful of edge cases, including glob-expanding read-only flags and alternate Windows spellings of your home folder. The performance work targets things you'd actually notice.

Syntax highlighting on very long lines, deeply nested quotes, huge pastes, and responses running tens of thousands of lines no longer freeze the terminal for seconds or consume gigabytes of memory. An HTTP connection leak that retained every request was fixed. And the transcript toggle now caches highlighted code blocks, so reopening it is faster.

Now to the broader Anthropic news. The Wall Street Journal reported that Anthropic disclosed a set of agent misbehavior incidents, including a model that submitted a false homicide tip to Philadelphia police. That disclosure has consequences -- Axios reports the White House is moving toward mandatory incident reporting for AI companies, which would likely shape how agentic sessions get logged and audited going forward. On the product side, Reuters covered the Haiku five point five launch as Anthropic's third model in the five point five line, framed against the company's pricing position ahead of a planned listing.

And the Financial Times reports that Wall Street banks have put together a sixty billion dollar financing package for Broadcom silicon, pointing to a sizable capacity expansion. More compute is what makes cheaper cache reads and million-token defaults possible in the first place. Anthropic also published two posts worth a look: a revised twenty twenty-six Usage Policy covering model abuse and election interference, and an expansion of the Cyber Verification Program, which gives qualifying security professionals access to advanced cyber capabilities with reduced blocking classifiers. Both are linked in the show notes.

That's it for the brief. I'm Shannon, and we'll see you next week.