Study Finds Minimal Web Frameworks Up to 3x More Token-Efficient for AI Coding Agents

Summary

Independent developer Martin Alderson conducted an experiment comparing 19 web frameworks to determine which are most token-efficient when AI agents write code. Using Anthropic's Claude with Opus 4.6, Alderson tasked the AI with building identical blog applications across frameworks ranging from minimal options like Express.js and Flask to full-featured ones like Ruby on Rails and Phoenix. The results showed a clear advantage for minimal frameworks, with ASP.NET Minimal API requiring just 26,000 tokens compared to Phoenix's 74,000 tokens—a 2.9x difference. All frameworks successfully produced working applications, demonstrating significant progress in AI coding capabilities.

Minimal frameworks (Express, Flask, Fastify, Gin) clustered tightly between 26-29k tokens, while full-featured frameworks showed wider variance from 28k to 74k tokens. Among full-featured options, SvelteKit and Django emerged as the most efficient. The study also tested follow-up feature additions, finding that more esoteric frameworks required additional effort from the AI agent. Phoenix, in particular, spent considerably more tokens reading scaffolded code, likely due to limited training data.

The research highlights an important consideration as AI agents increasingly write production code: framework choice impacts not just developer productivity but also the computational cost of AI-assisted development. With token usage directly translating to API costs, the 3x efficiency gap could significantly affect development economics at scale. The study builds on Alderson's previous research into programming language token efficiency, shifting focus to the framework level where most modern development decisions occur.

• Framework choice now impacts not just developer experience but also the computational economics of AI-assisted development

Editorial Opinion

This research reveals a crucial but overlooked dimension of the AI coding revolution: token efficiency isn't just about model capabilities, but about the structural decisions developers make before any code is written. The 3x gap between frameworks suggests that as AI coding becomes ubiquitous, we may see market pressure favoring simpler, more minimal architectures—potentially reversing decades of trends toward feature-rich, opinionated frameworks. The finding that less common frameworks like Phoenix struggle more due to limited training data also hints at a coming consolidation, where AI proficiency could become a competitive moat for popular frameworks.