These views are my own and do not reflect the views of any previous employer.

This week, Anthropic disabled token-based authentication for non-Anthropic apps. This was a long time coming with plenty of warning, but is still leaving a bitter taste in many devs' mouths.

I've been waiting for this

AI companies have been subsidizing token usage at an unsustainable rate for years. When agentic AI was new, you could use AI subscription tokens from Google (maybe OpenAI, too) with third party tools and it was clear to me at the time that this was a) Against TOS, b) Going to be expensive and patched, or c) both.

Google snubbed it pretty quickly. If you try to grab your API key from Google AI Studio, and use it with a third party app like Cline, you'll get hit with limits making it pretty much unusable. Link your account with Google Antigravity and use their agentic app, and you're in great shape.

Why they did it in the first place

AI companies benefit from subsidizing AI usage by creating user dependence on their tools. This makes sense for Claude Code. But if you can simply swap anthropic/opus-4.6 with model-du-jour in OpenClaw, there's not much upside for Anthropic to subsidize your AI usage if they're unlikely to retain your business anyway. A (probably Chinese) company will just blackhat train on Opus 4.6 and release the same thing for cheaper in a year anyways.

If you're using OpenClaw with the Claude subscription at an extensive capacity, you probably have the $100 or $200 subscription. This feels like a significant expense, and it is, but we must remember this is still a subsidized rate. $200/month is a lot, but it doesn't entitle you to unlimited free compute. The only reason you're even so hellbent on using your Claude token with OpenClaw/Codex/Whatever-tool is because you tried using these tools at-cost already, and realized how expensive AI tooling actually is.

We've been in a golden age for enthusiasts and people trying to build things quickly. Tech companies take a risk trying to be "the big one" by giving us free/cheap compute. We can take advantage (and have!) of free/cheap compute to create value.

Consider Amazon's long-term growth strategy: Delight customers at a defecit to earn trust, figure out the cashflow later. Anecdotally, Amazon Retail has squeezed customers considerably more in the past five years than the decade prior, Prime shipping is hit-and-miss, review trust is still terrible. None of this matters, though, because they invested in building dependence on immediate access to online goods. AI companies are investing in building dependence on AI tools.

The End of an Era is not The End of the World

People have been wondering if we're in an AI bubble. Most software engineers will tell you yes. Claude Code was the last hyper-generous API key for open usage. Its retirement marks the end of a gold-rush era. No more free pickaxes, I guess.

I got the idea for the title of this blog post writing the outro of this other post about automating SEO metadata and my dev workflow, you should check it out. If you have thoughts on AI, or feedback for a new blog author, tweet me at @natalieflora_ or email me at nfwalls@pm.me.

---

Published on April 6, 2026