April 18, 2026/838 words/4 min read

Stack Check, April '26

Claude got expensive, so the stack moved again.

I wrote Stack Check, March '26 six weeks ago and said Claude Code was the center of gravity in my setup.

Six weeks! Apparently that is how long it takes for an entire tooling stack to change these days. I would love to blame discipline, but no, this one is mostly Anthropic's fault. Up until recently, Claude made the rest of the stack feel calm and competent. It handled bigger async passes in the terminal without fighting the projects I gave it. I stayed in VS Code with Copilot for close editing and reviewed everything before it shipped.

But there we go, the stack has moved again, mostly because Claude has gotten really dumb. I feel like I am overpaying by a lot, even with the thousands of dollars of subsidized inference Anthropic throws at subscription users. Anthropic seems so terrified of Claude being used for malware that they half-lobotomized it with the system prompt, and normal coding work gets caught in the blast radius.

Claude Opus 4.7 especially made the annoyance concrete. Anthropic calls it a direct upgrade to 4.6 for difficult, long-running coding work, at the same price. Nice on the page! In practice, the sticker price stayed still while the effective cost moved underneath me. Exactly the "clever" bullshit that makes me want to go outside and touch grass instead of trying to build things.

The tokenizer analysis that made the rounds on HN measured the new tokenizer against real Claude Code content and found English and code getting meaningfully more expensive in token terms. The rough impact is a 20-30% worse session budget for the same work. When you use these tools as much as I do, that matters! I pay a little over 200 EUR a month for Claude, and around the same for ChatGPT. For the money, one of them is not like the other. At this point I probably will not even renew Claude, which would have sounded absurd a month ago.

The HN threads around 4.7 are full of people trying to work out whether the model improved at all. I do not buy the "more thoughtful" reading. It spends more tokens and feels less predictable, and that wears me down. A compiler does not wake up with a new personality when the vendor likes a new set of evals. My coding agent should not either!

That is also why VS Code and Copilot slid out of the default path. In March I wrote about Supermaven going away and how much I missed it, and that feeling has only sharpened. Cursor's completion spoiled me. It was fast enough to stay out of my way and really good at the small continuation work that makes an editor feel helpful. Copilot in VS Code just doesn't hit that mark. It is fine, but fine is the problem! VS Code used to win me back on performance, but lately it feels sloppier, and Copilot shows up everywhere.

So I moved to Zed.

I can disable AI in the editor entirely! Turning it off gets the Copilot chat sidebar and ghost text out of my way. I also lose tab autocomplete, which I miss quite a bit. I really had not realised how much of my typing rhythm had become "write half the boring thing, hit tab, keep moving". Losing that is a bit annoying, but at least the editor is just the editor again. I'm not sure if Zed's tab completion will do the trick, as I haven't tested it yet, but I'm thinking about it. Until then, the split feels healthier. I touch the code in Zed and give Codex the bigger pieces of engineering work that require spelunking through the repository. It can run the tests and bring back a diff for me to review.

And Codex has been a pleasant surprise! It understands the repository well enough to use the tools it finds there. It also explains what it is doing and generally behaves more like a pragmatic coding agent than a personality attached to a token furnace. That is the part of AI coding I still believe in.

Claude still has moments where it feels brilliant, but I no longer trust the product around the brilliance. The quota math stinks, and so does the vibes-based "better reasoning" story around the model changes. My own clients at work have taught me the hard way that trust is cumulative and fragile. Anthropic has spent too much of mine.

And still, annoyingly, I am optimistic!

When a tool understands the code well enough to plan the work and hand back a checked diff, it feels like the future I wanted. I do not want every surface in my working life to become a token furnace, though. For now I edit in Zed, which I can leave deliberately boring, and send the bigger work to Codex. Claude sits at the edge until I burn through what's left of the subscription. It is still useful, but I no longer trust it as the default.