MirrorCode

Coding unit: % independent 6 models scored not in composite

Code tasks built to mirror the structure of well-known problems while changing the underlying semantics, so memorised solutions fail.

What it measures

Whether coding ability generalises beyond memorised patterns.

Few models evaluated. Treat as directional.

#	Model	Lab	Score	Measured	Method	Source
1	Claude Fable 5	Anthropic	63.9%	2026-08-10	independent	Epoch AI Benchmarking Hub
2	Claude Opus 4.7	Anthropic	31.1%	2026-08-12	independent	Epoch AI Benchmarking Hub
3	GPT-5.6 Sol	OpenAI	20.0%	2026-08-10	independent	Epoch AI Benchmarking Hub
4	GPT-5.4	OpenAI	15.6%	2026-08-10	independent	Epoch AI Benchmarking Hub
5	GPT-5.5	OpenAI	10.0%	2026-08-10	independent	Epoch AI Benchmarking Hub
6	Gemini 3.1 Pro	Google DeepMind	8.9%	2026-08-10	independent	Epoch AI Benchmarking Hub