50x. That is the token reduction claimed by code-review-graph, a GitHub project that changes how code gets sent to AI models during code review sessions.
Tokens are the basic units AI models process - roughly 750 words per 1,000 tokens, and API costs scale directly with how many you use. When you paste code into Claude or GPT-4 for review, you are typically sending entire files, including all the functions, imports, and context the model does not actually need. For a 2,000-line file where 40 lines changed, you are paying for 1,960 lines of irrelevant text.
Code-review-graph takes a different approach: instead of dumping full files into the model's context window (the maximum amount of text a model can process in one session), it builds a dependency graph - mapping which functions call which, what depends on what - and sends only the relevant parts. You changed a function? The model sees that function, the functions it calls, and the functions that call it. Nothing more.
Does 50x Actually Hold Up?
The concept is sound. The number is almost certainly optimistic outside of ideal conditions.
A 50x reduction means a review that costs 50,000 tokens drops to 1,000. At Claude Sonnet's input pricing of around $3 per million tokens, that shifts a $0.15 review to $0.003. Across hundreds of daily reviews on a large engineering team, that is real money.
But the savings depend entirely on your codebase structure. A modular codebase where changes touch isolated functions would see large reductions. A legacy codebase where functions reference half the project will not see 50x - the dependency graph itself grows large, and the efficiency gains shrink accordingly.
The project is open-source and early-stage. If you are running AI code review at volume and paying meaningful API costs, it is worth testing on your actual code. Measure your token counts before and after on a representative sample, not on a best-case example picked to make the numbers look good. The technique has real merit. The specific multiplier needs real-world validation.