Code Review Eats 59.4% of Your Agentic AI Token Budget — Not Code Generation
The conventional assumption behind AI coding cost models is wrong. Engineers and finance teams have been tracking token spend on the assumption that the expensive part of agentic software development is code generation — the synthesis step. A January 2026 study from Concordia University-affiliated researchers, just surfacing on Hacker News, shows the actual breakdown is dramatically different.
Code Review accounts for 59.4% of all tokens in a typical agentic software engineering pipeline. Code generation itself is a rounding error by comparison.
What the Study Measured
The paper — “Tokenomics: Quantifying Where Tokens Are Used in Agentic Software Engineering” (arXiv:2601.14470) — analyzed 30 software development tasks executed by the ChatDev multi-agent framework using a GPT-5 reasoning model. Six distinct development stages were mapped and measured:
- Design
- Coding
- Code Completion
- Code Review — 59.4% of tokens
- Testing
- Documentation
Input tokens constitute 53.9% of total consumption across all stages — the largest single category. This reflects a structural property of multi-agent systems: agents repeatedly re-inject prior outputs as context for review and refinement passes. Each review cycle ingests the full accumulated context of the task so far.
The study concludes that “the primary cost of agentic software engineering lies not in initial code generation but in automated refinement and verification.”
Why This Changes the Cost Model
Most token budget tools and cost projections for AI coding agents are calibrated around output tokens — because output tokens have historically been the expensive side of the API ledger. They are also what you intuitively associate with a system “doing work.”
The Concordia findings invert that. The expensive operation is reading-plus-reviewing: each iteration of code review re-ingests an ever-growing context window of code, requirements, prior feedback, and agent state. A six-pass review cycle on a 2,000-line codebase doesn’t cost 6x the initial generation — it costs significantly more, because each pass processes the entire accumulated artifact.
Industry Context
The timing of this paper surfacing matters. GitHub Copilot switched from flat-rate to token-based billing on June 1. Enterprises across the industry are scrambling to reconcile AI tool budgets that arrived as surprises. TechCrunch reported this week that companies “gorged themselves on all-you-can-eat subscriptions” and are now “pulling back spending.” At least one hyperscaler customer reportedly spent $500 million in a single month after failing to set spend limits.
If code review is the dominant cost driver — not generation — then the optimization target shifts. Reducing review iteration count, compressing context windows mid-review, or routing lower-stakes review to smaller models are all higher-leverage interventions than optimizing prompt length at the generation step.
Implications for Agent Design
The findings point to a specific architectural opportunity: token-efficient review protocols. Rather than full-context re-injection on every review pass, agents could maintain differential context (tracking only changes since the last review), or apply hierarchical review (coarse-grained pass first, fine-grained only on flagged sections). The paper doesn’t propose solutions, but its empirical breakdown gives practitioners a concrete target.
The 53.9% share of input tokens also suggests that context compression — reducing what gets passed into each review step without degrading quality — could cut costs nearly in half without touching model capability.
Key Numbers
| Stage | Token Share |
|---|---|
| Code Review | 59.4% |
| All other stages combined | 40.6% |
| Input token share (total) | 53.9% |
Source: Salim et al., “Tokenomics: Quantifying Where Tokens Are Used in Agentic Software Engineering,” arXiv:2601.14470, January 20, 2026.