Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
I’m done.
TL;DR
Anthropic finally named the policy — and Theo thinks it’s a 25–40x rug pull — starting June 15, Claude paid plans get a separate monthly credit for programmatic use ($20/$100/$200 depending on tier), but Theo says that replaces what used to feel like $5,000–$7,500/month of effective Claude Code usage on the $200 plan.
The real split is now 'Anthropic UI' vs everything else — if you use Claude in Claude’s own chat, terminal, or desktop app, your subscription works as before, but if you use wrappers like T3 Code, Zed integrations, Claude-P, or the Agent SDK, you’re shoved into a tiny capped credit bucket instead.
Theo’s core complaint isn’t just pricing — it’s that Anthropic pushed developers onto the Agent SDK path and then changed the deal — he says tools like T3 Code were built around the closed-source Agent SDK specifically because Anthropic had repeatedly signaled that local, personal software built on it was allowed.
There is one genuinely good idea buried in the mess — Theo likes the concept of every Claude subscription including a small API-like experimentation budget, because it could help people prototype CI bots, audits, or PR workflows without needing a separate API account.
Anthropic’s business logic makes sense to him; the implementation doesn’t — he argues they clearly want to stop token-hungry tools like OpenClaw and other automations from burning subsidized compute, but says they drew the line too early and punished benign UI wrappers along with real 'token farms.'
Theo ends by canceling his Claude plan and telling users to look elsewhere — after saying Anthropic’s software is unreliable and hostile to open-source builders, he points people toward alternatives in T3 Code like Codex, Cursor, OpenCode, Gemini, and GitHub Copilot support.
The Breakdown
Anthropic finally answers — and Theo instantly hates the answer
Theo opens wearing a Claude hat because Anthropic finally responded to months of confusion around Claude Code subscriptions. The headline: paid Claude plans can claim a monthly programmatic-use credit starting June 15, and Anthropic’s PM explicitly cites tools like T3 Code and OpenClaw — which makes Theo say, yes, this absolutely targets him.
The policy sounds generous until you read the fine print
He zeroes in on the wording: users must manually "claim" the credit, it doesn’t roll over, and once it’s gone you fall back to API-priced usage credits. On paper, $20/$100/$200 back for programmatic usage sounds nice; in Theo’s framing, it’s wildly misleading because it doesn’t preserve the huge effective subsidy people were getting through normal Claude Code subscription usage.
Why this was always a messy spectrum, not a clean rule
Theo lays out the usage ladder: Claude chat on the site, then Claude Code, then tools like OpenCode, then raw API calls. His point is that Anthropic has decreasing control as you move rightward, and that matters because a $200 Claude Code plan can effectively unlock thousands in inference — he cites estimates of $5,000, maybe even $7,500/month after recent limit increases — while enterprise API customers pay something much closer to full freight.
The hidden business model: consumer subsidy, enterprise monetization
This is Theo’s bigger theory of the case: Anthropic doesn’t subsidize heavy personal users out of generosity, it does it because those users bring their habits to work. He says engineers can burn $10k/month in enterprise inference, and that’s part of why Anthropic’s reported revenue has recently surpassed OpenAI’s — not because casual users under-consume, but because enterprise buyers absorb the real cost.
OpenCode, caching drama, and how the lockout started
Theo argues Anthropic got stricter after people used Claude Code subscriptions inside OpenCode, a better harness in his view than Claude Code itself. He brings up Dax’s tests on tool-call pruning and caching to argue Anthropic’s own caching approach is flawed, then says OpenCode users likely cost Anthropic more while giving them less product lock-in — the exact combo that would motivate a crackdown.
The Agent SDK trap: the path Anthropic told developers to use
The most personal section is about the Agent SDK and Claude-P. Theo says Anthropic repeatedly implied these were valid for local development and experimentation, which is why people like Matt Pocock, José Valim, Zed, Gene, and T3 Code built on them while waiting for formal clarity that never came.
The line Anthropic drew is the one Theo can’t forgive
Theo says Anthropic should have separated UI wrappers from pure automation tools like OpenClaw or Hermes Agent, which can constantly burn tokens. Instead, they grouped nearly all non-Anthropic interfaces into "programmatic" use, so something human-in-the-loop like T3 Code gets treated closer to a cron-job token farm than to Claude Code itself.
What breaks now — and why Theo calls it an attack on open source
He runs through the casualties: T3 Code, Zed’s Claude ACP adapter, CI integrations, PR bots, security audit workflows like Project GlassWing, even calling claude -p from your own terminal. His blunt summary is that if you can’t see the Claude logo and you are not inside Anthropic’s closed-source UI, you’ve been cut off from subscription economics.
Theo’s workaround, final verdict, and breakup text to Anthropic
Because of the change, Theo says T3 Code will likely ship an option that uses Anthropic’s own terminal UI just to preserve higher subscription limits — a workaround he clearly resents. He ends by canceling his Claude plan, quoting Matt Pocock’s "poisoned chalice" line, and saying the final lesson is simple: Anthropic told builders to trust this path, then pulled the rug anyway.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
