HelpBench: leaderboard

Metric: Overall rubric score (%): points for met positive criteria and avoided negative criteria (factual criteria written per question by privacy, safety and security experts, plus shared delivery criteria) over the maximum points, applied by a Gemini 2.5 Pro auto-rater at temperature 0; chat versions with default parameters, mean of five responses per question, 450 questions; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 18 models tracked.

Top models

#ModelScore
1GPT-5.387
2GPT-5 Chat87
3GPT-4.184
4Claude Opus 4.684
5Gemini 3 Flash84
6Claude Sonnet 4.683
7Grok 483
8Qwen 3.6 Plus83
9Gemini 3.1 Pro (Preview)82
10Claude Sonnet 482
11GLM-5.182
12Gemini 2.5 Pro81
13GLM-4.680
14Grok 4.2080
15DeepSeek V3.2 (Non-reasoning)80

Interactive version: theaggregate.ai/benchmark?slug=helpbench · How It Works · Data refreshed daily, snapshot 2026-09-29.