Anthropic announced Code Review for Claude Code earlier this week. The internet immediately fixated on the price — $15-$25 per review — and whether that’s a good deal relative to a senior engineer’s time, but they missed the interesting part.

Anthropic sells tokens like tire companies sell tires. Code Review is less a boxed copy of Microsoft Word 2000 on a shelf and more an entry in a Michelin Guide — a curated experience that exists to put miles on rubber. The $15-$25 is the cost of driving somewhere.

Any alternative a buyer comes up with still consumes tokens. You can roll your own review prompt, or use a competitor’s wrapper, but the meter is always running on somebody’s foundation model. Anthropic doesn’t need to lock you into their review product specifically - they just need you reviewing code with LLMs at all. That they don’t mark up the tokens that this product consumes means they’re hard to beat by anybody other than another foundation model.

I’m not sure “product” is even the right word for what this is. It’s a complex prompt with good defaults — closer to a recipe than a piece of software. You can modify it, replace parts of it, ignore it entirely and do the same thing yourself. Richard Stallman spent decades arguing that software should work kind of like this. Turns out the economics get a lot simpler when the vendor’s business model doesn’t depend on the software itself, just on the fuel it burns. The tires still get sold.

I expect Anthropic and every other foundation model company to repeat this exercise: ship opinionated workflows that look like products but function as on-ramps for token consumption. Some will be genuinely great, and all of them will sell tires.