Announcing gmickel-bench: Real-World Evals

Been wanting to do this for a while.

Most AI benchmarks tell you if a model can solve a LeetCode puzzle. They don't tell you if it can ship a product.

I got tired of wondering which model would actually help me build real things. So I built gmickel-bench: a living scoreboard based on the real-world engineering tasks I actually ship.

What I Test

Three domains. Three very different skill sets:

• Full-Stack: Building OAuth 2.1 servers on Convex
• Frontend: Designing complex Next.js dashboards from scratch
• Systems: Writing low-level memory management code in Zig

The Models

I ran three models through the gauntlet:

• Claude Code with Opus 4.5
• Gemini CLI with Gemini 3.0 Pro
• OpenAI Codex with gpt-5.1-codex-max medium

What I Found

Claude Code: The Product Architect

Best overall. Crushes frontend design and auth flows. Struggles when the math gets low-level.

Gemini CLI: The Systems Hacker

The only model that survived the Zig torture test (build → train → sample) without crashing. Less polish on the product side.

OpenAI Codex: The Pattern Matcher

Good at refactoring. Good at isolated pages. Feels lazy when you need comprehensive implementations.

See the Data

The full benchmark, scoring methodology, and detailed breakdown is live at mickel.tech/gmickel-bench.