How well do AI models meet education-relevant constraints?

ELBench scores AI models on 2,481 questions in four areas: teaching, safety, knowledge & reasoning, and character & thinking skills. Each dot is a model: higher scores better, further left costs less.

Include the region-specific sets (the paper's original scoring)

Adds two sets that test region-specific content rules and national exams. See About.

Graded by

Each judge model has its own leaderboard. The published results were graded by the paper's judge, so they appear on the default one.

Highest price

Dollars per million tokens. All the way right shows every price.

0.1 30

Largest model size

Billions of parameters. All the way right shows every size.

1 1000

Include models that don't publish their size

Most closed models, such as GPT, Claude and Gemini, don't.

Best value: DeepSeek-V4-Flash (82.8 at $0.11) → Gemini-3.5-Flash (85.6 at $3.38) → GPT-5.4 (86.3 at $5.62). Each of these scores higher than every model that costs less.

Not on OpenRouter, so not plotted: Doubao-Seed-2.0-Pro, Safe-InnoSpark, InnoSpark-235B.

_{Scored on 2,481 questions, without the two region-specific sets (see About). Published 95% ranges cover the paper's original scoring only, so none are shown here. Prices from OpenRouter, 26 Oct 2026 23:21 UTC.}