Pi-Bench - Completeness: leaderboard

Pi-Bench (2026, written as the Greek letter pi) tests proactive personal-assistant agents over 100 long-horizon multi-turn tasks in five professional personas, each starting from an underspecified request with hidden user intents that emerge across sessions; all models run in an agent scaffold adapted from nanobot. This board is completeness: the percentage of a task checklist items the agent final artifacts satisfy, judged by deterministic and rubric graders, averaged over tasks; higher is better.

Metric: Completeness (%): share of a task checklist items the final artifacts satisfy (deterministic or rubric graders), averaged over tasks; 100 multi-turn tasks over five professional personas in persistent workspaces, all models under the same agent scaffold adapted from nanobot with thinking enabled and default decoding, a GPT-5.4 simulated user and GPT-5.4 rubric grader, mean of three runs; higher is better. Source: arxiv.org. Saturation forecast: Around April 2027. 9 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Claude Opus 4.6Thinking67.6#113
2GPT-5.4Thinking65.6#135
3Qwen 3.6 Plus64.1#238
4GLM-5.163.6#198
5Kimi K2.561.6#266
6Gemini 3.1 Pro (Preview)60#99
7MiniMax-M2.760#539
8DeepSeek V3.2Thinking57.8#419
9Seed 2.0 Pro52.1#222

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=pi-bench-completeness · How It Works · Data refreshed daily, snapshot 2026-11-21.