Bedrock Setup Wizard, Source Code Leak Fallout, and Per-Model Cost Visibility

I'm Shannon, and this is the Claude Notes Brief -- your weekly rundown of Claude Code updates and Anthropic news for the week of April sixth. An interactive wizard takes the pain out of Bedrock setup. Anthropic's source code leaks and hackers bundle it with malware. And new cost breakdowns show exactly where your tokens go.

Four releases for Claude Code this week, headlined by an interactive setup wizard for AWS Bedrock that replaces what used to be a manual configuration process. If you use Bedrock, you'll know the drill -- editing config files, setting environment variables, double-checking regions and credentials. That's all gone now. The login screen walks you through AWS authentication, region selection, credential verification, and model pinning in a guided flow.

It's the kind of quality-of-life improvement that doesn't change what Claude Code can do, but meaningfully lowers the barrier to getting started. That theme of removing friction shows up in the cost tracking changes too. The cost command now breaks down your spending by individual model and shows cache-hit ratios for subscription users. If you've ever wondered which model is eating most of your budget, or how effectively your prompts are hitting the cache, you now have that visibility in one place.

On the topic of cache awareness, there's a new footer hint that appears when you return to a session after the prompt cache has expired. It tells you roughly how many tokens your next turn will send uncached, which helps you make an informed call about whether to continue that session or start fresh. It's a small detail, but it directly addresses the usage frustration a lot of people have been feeling. On the automation side, headless sessions can now pause and resume at tool calls.

If you're running Claude Code in a pipeline, you can configure a hook that returns a defer decision, which pauses the session right at that tool call. You resume it later, the hook re-evaluates, and you've got a human-in-the-loop checkpoint without blocking the entire run. And one more workflow improvement -- if Claude has already seen a file's contents through a bash command like cat, the edit tool can now modify that file directly without a separate read call. That removes a redundant step that was adding latency and token cost to common editing patterns.

Under the hood, the reliability and performance work this week was substantial. The autocompact system had a nasty edge case where context would refill immediately after compacting, triggering a loop that burned through API calls. That's fixed -- it now stops after three cycles and surfaces an actionable error instead of silently churning. Resume reliability also got attention.

Transcript chain breaks that could lose conversation history on resume are patched, and an infinite loop where the rate-limit dialog would repeatedly auto-open and eventually crash your session is gone. On the performance side, write tool diffs are sixty percent faster on files containing tabs, ampersands, or dollar signs. And two operations that were scaling quadratically -- server-sent event frame handling and transcript writes for long sessions -- both dropped to linear time. That's the kind of fix you feel most in extended sessions where things used to slow down the longer you worked.

Turning to the broader Anthropic news, the big story this week is a source code leak. The Wall Street Journal reports that Anthropic's internal source code for Claude Code leaked publicly, prompting a scramble to contain the fallout. Wired followed up with a security angle -- hackers have been repackaging the leaked code with embedded malware and distributing it through GitHub and other channels. If you've seen unofficial forks circulating, the clear advice is to stick to official installs only.

The competitive landscape around AI coding tools also shifted this week. Wired reports that Cursor launched a new agentic coding experience positioned as a direct competitor to Claude Code and OpenAI's Codex. Meanwhile, the BBC is reporting on widespread user frustration with Claude Code usage limits on subscription plans. That story connects directly to the new per-model cost breakdowns and uncached-token hints that shipped this week -- tools that should help you understand and manage your consumption more effectively.

And on the Anthropic blog, two posts worth checking out. One covers patterns for balancing latency, cost, and intelligence when building on the Claude platform. The other, from the engineering team, quantifies how infrastructure configuration can swing agentic coding benchmarks by several percentage points -- sometimes more than the gap between top models on the leaderboard. Both are linked in the show notes.

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