PluginEval: leaderboard

Metric: Query-level accuracy (%; correct tool calls with correct arguments on positive queries and abstention on negative queries; original difficulty distribution; 3,000 human-verified Chinese function-calling queries over 54 plugins and 63 tools with real API execution (2,500 positive queries requiring grounded calls, 500 adversarial negatives requiring abstention; easy/medium/hard 0.4/0.4/0.2); GPT-5.2 judge anchored to gold annotations). Source: arxiv.org. Saturation forecast: Around August 2027. 5 models tracked.

Top models

#ModelScore
1Claude Opus 4.668.17
2Gemini 3.1 Pro (Preview)67.78
3DeepSeek V4 Pro64.33
4Qwen 3 235B A22B58.4
5GPT-5.452.13

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