Claude Code v2.1.296 Just Halved Every Agent Bill

Show Notes

Watch the video: https://www.youtube.com/watch?v=jwTCtikSL2Y

Anthropic shipped Claude Code v2.1.296 yesterday and quietly halved the Sonnet 5.5 cache read price - the single biggest line on any agent bill that re-reads your codebase daily.
Playbooks + my answers, free - Church of SEO Jesus: https://www.skool.com/church-of-seo-jesus

What you get in 7 minutes:
- The one env var (CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL) to send your developer today
- Why Sonnet 5.5 cache reads dropping from $0.20 to $0.10 per million tokens cuts your weekly agent invoice roughly in half
- The two Windows fixes that stopped silent file corruption on Shift-JIS, GBK and Windows-1252 codebases
- The PowerShell allow-rule ceiling jumping 32x (from ~1 KB to 32 KB)
- The MCP tool description cap doubling to 4,096 characters

Sources:
Claude Code v2.1.296 release notes: https://github.com/anthropics/claude-code/releases/tag/v2.1.296
Anthropic pricing: https://www.anthropic.com/pricing

Chapters below.

Chapters:
00:00 Claude Code v2.1.296 shipped yesterday
00:22 Operator take on v2.1.296
00:45 Sonnet 5.5 cache reads halved to $0.10/M
01:18 Who actually hits the cache
01:45 CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL explained
02:21 Subscribe + Free Church ask
02:52 Terminal demo on the release notes
03:55 Windows PowerShell 32 KB fix
04:25 Edit tool refuses Shift-JIS and GBK files
05:01 allow_large, autoCompactWindow, 4,096 MCP cap
05:45 Who gets undercut this week
06:05 Church of SEO Jesus playbook
06:30 One action in the next ten minutes
06:56 Haiku vs Opus bill benchmark

Audio edition of the Stewart Vickers AI YouTube video of the same name. Full transcript on this episode page.

What is Stewart Vickers AI?

The audio edition of the Stewart Vickers AI YouTube channel. Each episode takes one AI release - a model, a tool, a price cut - and translates it into money for business owners: what got cheaper, who got undercut, and what to do about it this week. Operator credibility over theory, specific numbers over vibes. Every episode links the original video and the free Church of SEO Jesus community where the prompts and workflows live.

Anthropic halved the price yesterday. 296 shipped with one line that cut the ongoing cost of every agent you have running against the same codebase. Nothing on your bill alerted you. Open the release page yourself and the engineers are the only people in the comments. Here is the operator take. One env via your developer flips today and Opus stops paying Opus prices for grunt work on your workflows.

One price change halves your cash bill. Two Windows fixes stop silent file corruption. I will show you the real command, the real output, and the single setting to send your dev before you close this tab. Start with the money line. Sonnet 5. Five cash reads just dropped from 20 cents PR million tokens to 10 cents PR million tokens.

Cash reads are what an agent pays every time it re-reads the same repo, the same docs, the same brand brief. If you run a Claude Code agent on your codebase every day, the biggest line on your invoice is the one that keeps paying to re-read what it already read yesterday. And Anthropic halved that line. Zero config. Update your Claude Code and it applies. The caveat matters.

Opus and Haiku are not repriced here. Only Sonnet 5. Five. The savings only land if your workflow actually hits the cash. A one-shot run against a fresh repo barely moves. An agent that lives in your codebase and re-reads your knowledgebase every turn moves a lot.

I run one that re-reads a 600-page client brief every pass. The weekly bill on that agent falls roughly in half starting Monday. Here is the one setting to send your developer in Slack right now. 5. What it does, your workflow fans out dozens of throwaway sub-agents to do grunt work, grep this folder, summarise that doc, check a status, and before yesterday every one of them ran on whatever flagship Modal the parent chose. Opus prices for Opus-grade work on the main loop.

Opus prices for a glorified grep inside it. Now the main loop stays on Opus and the fan-out runs on Haiku. The savings on a workflow-heavy pipeline are embarrassing. Subscribe to this channel because every Claude code release gets translated into money the day it ships, so the next drop finds you instead of you finding it on an invoice 4 weeks later. The exact commands I ran, the env of our name, and the two windows fixes are in the pinned comment below. And the Church of S-E-O-Jesus, my free school community, is where the playbook versions of these videos live and where I answer when your setup breaks.

Join link in the description, free, come in tonight. Proof. I ran this during the recording. Pull the release notes with invokeRest method, grep the body for the 5 lines that matter, and the first 6 results come back clean. Workflow sub-agent Modal, allow large on the read tool, Windows PowerShell commands over 1KB stop asking for permission, the slash cost status line, updated to the new 10 cent cash price. The exact text sits on the GitHub page.

Then I pipe the same 5 changes into my own Claude helper and ask it for a 1 sentence operator take at opus 5, the answer comes back in 126 output tokens. Tell your developer to set workflow underscore sub-agent underscore Modal to haiku today, because workflow runs fan out dozens of throwaway sub-agents whose grunt work does not need a flagship Modal, so pinning them cheap cuts the biggest line on your bill with no loss on the parts that matter. The agent and I arrived at the same conclusion by the same route, the release is not subtle, it is simply unread. Now the Windows half, which almost nobody is covering. If your dev team is on Windows, and your allow list exists for a reason, Claude Code was asking permission on every PowerShell command over about 1KB, 1K is nothing. One decent pipeline through 5 chem blitz blew past that ceiling.

Yesterday the ceiling moved to 32KB. Allow rules you already wrote work on realistic commands now, your developer gets back the time they used to spend tapping Y. The second Windows fix is quieter and scarier. If your codebase holds files that are not valid you tf8, windows 1252 config, shiftjis from a Japanese library, gbk from a Chinese partner, Claude Code's edit tool was silently replacing every non-ASCII character in them with a question mark. The German umayorps went. The Japanese katakana went.

The Chinese product names went. Nothing failed. The commit went in and the file was wrong. Yesterday the tool refuses those files outright instead of corrupting them. If you have ever shipped a localization bug to a Japanese customer you now know why. Three smaller operator wins.

Allow large on the read tool. Claude can swallow a full file past the old size cap in one call. A 3, 0, 0, 0 line log, a long transcript, a legal doc, before it read a slice and guessed. Auto-compact window on subagents. Long overnight jobs used to die when one subagent ballooned its context and dragged the parent down. Now each subagent compacts itself earlier and your pipeline wakes up still alive.

MCP tool descriptions doubled from 2,048 to 4,096 characters. If you connect your own CRM, your email, your stripe to Claude, the description Claude reads to understand what the tool does just doubled. Your tools can be richer. Who gets undercut this week? Agencies billing you flat for an AI build. Anybody charging you a PR token markup.

The dev shop that quoted you a Claude agent build last month and has not updated their pricing. Your competitor who is reading these notes and flipping that envar tonight is going to ship the next build at a lower marginal cost than the agency bidding against them. The window is between now and whenever your competitor sees this video or reads the changelog themselves. A week, maybe two. The Church of SEO Jesus, my free school community, is where I post the exact envar strings and the measured build drops from my own build each week. 296 lives in there.

Every release note I translate into money. I post the companion checklist first for the members. And when your dev team's Claude code setup breaks on an edge case I did not cover tonight, that is where I answer. Join link is pinned in the description. Free. Come in tonight.

Open the release page yourself right now. Search for cash reads. The 10 cent line is the proof, not me telling you. One thing to do in the next 10 minutes. Open Slack. Message your developer or your agency.

Paste the envar name. Send them this video. Ask them when they can update your Claude code. And flip the setting. If they say they already have, ask them to show you the slash cost line and confirm the cash read price now reads 10 cents. The number is right there in the status line.

The invoice drop lands by the end of the week. The Haiku workflow I just told you to flip. I benchmarked it against the opus-only build last night. Same prompt, same repo, two builds side by side. And the end screen is where that comparison lives.