LLM Racing Games pits models head-to-head

This is the kind of comparison that’s valuable precisely because it’s imperfect: it exposes not just output quality, but how models behave under iterative, tool-using coding workflows.

• –The results read like a qualitative benchmark for agentic coding, not a strict eval, which makes the differences more interesting than a simple score table.
• –The post highlights distinct failure modes: regressions, overlong code dumps, broken tool setups, invisible track logic, and one model that only improved after Playwright MCP was accidentally disabled.
• –The strongest signal is variance in execution style, not just end-state polish: some models edited incrementally, others rewrote everything, and some leaned into hidden structure or side effects.
• –It also shows how much the evaluation setup matters. Vision, browser tooling, and prompt iteration all materially changed outcomes, so apples-to-apples comparisons are only partly achievable.
• –As a shareable artifact, it’s compelling because people can play the demos themselves and judge the tradeoffs directly rather than trusting a static leaderboard.