The Models Engineering Index (MEI) compares language models on practical engineering work, combining results across benchmarks that include CursorBench, FrontierCode, Migration, IOI, ProgramBench, Terminal-Bench, and Vibe Code. Google’s 4 Neon ranks first with an index score of 69.5, ahead of Anthropic’s Fable 5.2 at 65.8 and Opus 5.5 at 63.1. OpenAI’s 6.2 Astra places fourth at 60.5, while Google’s 4 Hydrogen and Anthropic’s Sonnet 5.5 tie at 57.2. The index is intended to show performance on engineering tasks rather than provide a general-purpose ranking of language models.
The rankings also present cost and token use as separate considerations, making clear that a higher score can come with substantially greater resource demands. Relative figures show Fable 5.2 at 6.5 times the listed cost and 2.1 times the tokens of the comparison baseline, versus 4 Neon at 1.9 times the cost and baseline-level token use. Astra scores below the top three but uses only 0.29 times as many tokens; 6.2 Sol scores 56.9 while listing cost and token use at 0.4 and 0.35 times. Lower-ranked models include inexpensive options such as 6 Luna and V4.1 Flash, whose scores are 34.8 and 44.1. The underlying data is attributed to Cognition, Cursor, the Laude Institute, and Vals AI.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.