Redditor makes 45 models draw the same car, agents self-correct via screenshots
qu1dow · reddit · 2026-11-20
Reddit user qu1dow ran a fun benchmark: 45 models (including quants) got the same prompt — "an animated side view of a car driving through a landscape, one HTML file."
- Models ran as agents with a browser, so they could screenshot their own output and fix it, forming a self-correction loop
- Lineup: Claude (Opus/Sonnet/Haiku 4.5–5.5), GPT-5.6 to 6.2, Qwen, DeepSeek, GLM and others, plus local Qwen3.8 quants
- All results viewable at car-bench.pages.dev; author cautions the ranks aren't rigorous
A spin on the classic "pelican on a bicycle" benchmark, offering an intuitive look at how models differ in frontend generation and agentic self-correction.
More from coding & agent
- Claude HQ now supports connecting two Notion workspaces at once — jacob_posel · 2026-11-20
- WebGPU pinball build runs live in browser: 480Hz physics, Blender parts, wet neon Tokyo alley — AI_Andrew · 2026-11-20
- Inside OpenAI's agent orchestrator: task service, dispatch loop, and event subscriptions — OpenAIDevs · 2026-11-20
- OpenAI: Codex regression workflow ran 1,700+ investigations, found root causes in 600+ cases — OpenAIDevs · 2026-11-20
- OpenAI's Codex-powered 'Performance Factory' has run 1,700+ regression investigations in 3 months — OpenAIDevs · 2026-11-20
- Gradio 7 will cut its dependencies in half over supply-chain security concerns — _akhaliq · 2026-11-20