Google's Gemini 4 Neon tops the Models Engineering Index (MEI) with a composite score of 68.1 across ten coding benchmarks, adjusted for difficulty and missing results against a 50-point mean. The index normalizes cost and token usage relative to a typical model baseline of 1×. Gemini 4 Neon achieves its lead with 1.6× the reference cost and 0.83× token consumption. Anthropic's Fable 5.2 ranks second at 66.3 but incurs significantly higher resource demands at 5.9× cost and 2.1× tokens. OpenAI's 6.2 Astra places third with 62.0, operating efficiently at 1.5× cost and 0.31× tokens. Anthropic's Opus 5.5 follows at 60.8 (3.6× cost, 2× tokens), while Google's 4 Carbon scores 60.6 (1.7× cost, 0.88× tokens). Sonnet 5.5 reaches 57.9 with 3× cost and 3.4× tokens. Google's 4 Hydrogen sits at 56.5 (1.1× cost, 1× tokens), matched closely by OpenAI's 6.2 Sol at 56.0 (0.36× cost, 0.31× tokens). Alibaba's 4-Max scores 49.2 at 7.1× cost. DeepSeek's V4.1 Flash achieves 44.5 with notably low cost at 0.12× but 1.5× tokens. xAI's Grok 4.7 scores 41.7 (3.4× cost, 0.8× tokens). OpenAI's 6 Luna trails at 34.4 with minimal cost (0.063×) and 0.64× tokens.

Read original