Code Review and the end of subsidized tokens

Yesterday, Anthropic launched Code Review, a tool built into Claude Code that sends multiple agents in parallel to read every pull request before it reaches human hands. The agents look for logic errors and security vulnerabilities, then cross-verify each other's findings to filter false positives and rank issues by severity. Results surface as inline comments directly on the GitHub PR. Technically it's interesting: more than just static analysis or style enforcement, it's repository-aware reasoning. Before deploying it internally, only 16% of Anthropic's own PRs received substantive feedback; with the tool active, that number jumped to 54%.

However, each review averages between $15 and $25 according to Anthropic themselves, scaling with PR size and complexity. That's PER pull request. FindArticles points out the very interesting example of a team of 100 developers opening one PR per workday, that's roughly 2000 PRs a month - about $40,000 at the midpoint price. And here's where it gets funny, because the full loop looks like this: Claude Code writes the code, introduces bugs, and then Code Review finds them and charges you $20 for the trouble.

But beyond the joke, there's something more worth naming. These $15-25 per review are real tokens, consumed by fleets of agents running in parallel for about 20 minutes per PR. It's not an inflated number, but what deep reasoning over code actually costs at scale. And that raises the, by now, classic question: how much longer can AI labs sustain the flat subscription pricing over workloads that are, intrinsically, expensive to run? Tokens have been subsidized, and products like Code Review and Perplexity Computer (and anything that comes in the future) are starting to show how much. Anthropic frames it as an insurance product - "the cost of a shipped bug dwarfs $20 per review", they told VentureBeat - and they're probably right on that argument. But the argument also runs in reverse: if this is what it actually costs to review a PR properly, then what we've been paying for "unlimited" on our $20/month plans was always a form of accounting fiction.

OK, the era of subsidized tokens probably doesn't end tomorrow, not probably this year even. But! Code Review is a signal of where the industry is pointing, toward pricing that reflects the real cost of inference, especially as workloads become agentic and complex. For enterprise teams already deep in Claude Code, $20 per PR might be perfectly reasonable. For everyone else, it's a preview of the bill that's been quietly accumulating behind the scenes.