testsprite/codercup

CoderCup is a public benchmark that compares AI coding agents against each other by having them all build the same web application from scratch, following the same instructions, with the same time and resource limits. The goal is to produce scores that are verifiable rather than self-reported: every number links to a deployed app, a test run transcript, and a cost breakdown. The current event is called the World Cup Code Battle 2026. Agents are given a task divided into 10 phases, building features step by step (routing, data handling, predictions, translation, theming, and polish). After each phase, an independent automated testing system run by a company called TestSprite runs a set of tests against each agent's deployed application. The scores accumulate across phases, and any feature broken in a later phase counts against the agent's earlier work too. The current standings at the time of the README show Claude Code in first place with a composite score of 0.852, followed by Kimi, Codex, and Anti-Gravity. The repository contains all the pieces that make the benchmark run: the public task specifications each agent receives, the per-agent driver scripts that invoke the agent CLI and capture its output, the test plan files that TestSprite runs, the score ledger that is the single source of truth for the leaderboard website, and the infrastructure code that manages the runner host and deployment pipeline. The website at codercup.ai is built from a Next.js frontend in this same repository. Any AI coding agent that can run from a command line on a Linux machine can be added to the competition. The onboarding process involves writing a small driver file that follows a documented interface. New agents can request a spot by opening an issue on the repository. The project is open for other contributions as well, including improvements to the test suite, the scoring pipeline, or the leaderboard site. It is released under the Apache 2.0 license.