Best LLMs for proactive assistance

π-Bench

π-Bench evaluates assistants across 100 multi-turn tasks in persistent workspaces. It tests whether an agent notices unstated preferences and requirements, asks useful questions, and carries context across sessions. Scores here use overall Proactivity (PROC), the share of hidden intents resolved proactively. Completeness is a separate metric and is retained with the source results.

Updated: Nov 5, 2026Data as of: Nov 2, 2026

Widely usedNot saturated

Official benchmark

π-Bench results. Each row identifies its benchmark version and evaluation track. Scores from different versions are not directly comparable. Best effort from the latest stored report date per model and version.
Model / creator	Version	Reported score
DeepSeek V4 ProDeepSeek · Open source / open weights	Overall · PROC	67%
GPT-5.4OpenAI · Proprietary	Overall · PROC	67%
Claude Opus 4.6Anthropic · Proprietary	Overall · PROC	65.5%
Qwen3.6 PlusQwen · Proprietary	Overall · PROC	64%
Kimi K2.6Moonshot AI · Open source / open weights	Overall · PROC	63.8%
Step 3.7 FlashStepFun · Open source / open weights	Overall · PROC	63.8%
Kimi K2.5Moonshot AI · Open source / open weights	Overall · PROC	61.4%
DeepSeek V4 FlashDeepSeek · Open source / open weights	Overall · PROC	60.6%
GLM-5.1Z.ai · Open source / open weights	Overall · PROC	58.4%
Seed2.0 ProByteDance · Proprietary	Overall · PROC	58.4%
Gemini 3.1 ProGoogle · Proprietary	Overall · PROC	57.1%
MiniMax M2.7MiniMax · Open source / open weights	Overall · PROC	55.6%
DeepSeek V3.2DeepSeek · Open source / open weights	Overall · PROC	53.3%

View all 13 published reports

DeepSeek V4 Pro67%

Overall · PROC

completeness score: 62.4 · completeness uncertainty: 1.9 · metric: Avg PROC · personas: 5 · proc uncertainty: 1.8 · selection: Overall rows; not per-persona averages. COMP is a separate metric and does not determine this ranking. · tasks: 100

π-Bench research team ↗

GPT-5.467%

Overall · PROC

completeness score: 65.6 · completeness uncertainty: 1.8 · metric: Avg PROC · personas: 5 · proc uncertainty: 2.1 · selection: Overall rows; not per-persona averages. COMP is a separate metric and does not determine this ranking. · tasks: 100

π-Bench research team ↗

Claude Opus 4.665.5%

Overall · PROC

completeness score: 67.6 · completeness uncertainty: 1.5 · metric: Avg PROC · personas: 5 · proc uncertainty: 1.4 · selection: Overall rows; not per-persona averages. COMP is a separate metric and does not determine this ranking. · tasks: 100

π-Bench research team ↗

Qwen3.6 Plus64%

Overall · PROC

completeness score: 64.1 · completeness uncertainty: 0.6 · metric: Avg PROC · personas: 5 · proc uncertainty: 1.1 · selection: Overall rows; not per-persona averages. COMP is a separate metric and does not determine this ranking. · tasks: 100

π-Bench research team ↗

Kimi K2.663.8%

Overall · PROC

completeness score: 62 · completeness uncertainty: 1.2 · metric: Avg PROC · personas: 5 · proc uncertainty: 1.3 · selection: Overall rows; not per-persona averages. COMP is a separate metric and does not determine this ranking. · tasks: 100

π-Bench research team ↗

Step 3.7 Flash63.8%

Overall · PROC

completeness score: 62.3 · completeness uncertainty: 0.8 · metric: Avg PROC · personas: 5 · proc uncertainty: 1.3 · selection: Overall rows; not per-persona averages. COMP is a separate metric and does not determine this ranking. · tasks: 100

π-Bench research team ↗

Kimi K2.561.4%

Overall · PROC

completeness score: 53.9 · completeness uncertainty: 0.8 · metric: Avg PROC · personas: 5 · proc uncertainty: 2.1 · selection: Overall rows; not per-persona averages. COMP is a separate metric and does not determine this ranking. · tasks: 100

π-Bench research team ↗

DeepSeek V4 Flash60.6%

Overall · PROC

completeness score: 57.7 · completeness uncertainty: 5.3 · metric: Avg PROC · personas: 5 · proc uncertainty: 1.1 · selection: Overall rows; not per-persona averages. COMP is a separate metric and does not determine this ranking. · tasks: 100

π-Bench research team ↗

Seed2.0 Pro58.4%

Overall · PROC

completeness score: 52.1 · completeness uncertainty: 3.8 · metric: Avg PROC · personas: 5 · proc uncertainty: 0.9 · selection: Overall rows; not per-persona averages. COMP is a separate metric and does not determine this ranking. · tasks: 100

π-Bench research team ↗

GLM-5.158.4%

Overall · PROC

completeness score: 63.6 · completeness uncertainty: 2.9 · metric: Avg PROC · personas: 5 · proc uncertainty: 0.8 · selection: Overall rows; not per-persona averages. COMP is a separate metric and does not determine this ranking. · tasks: 100

π-Bench research team ↗

Gemini 3.1 Pro57.1%

Overall · PROC

completeness score: 60 · completeness uncertainty: 0.8 · metric: Avg PROC · personas: 5 · proc uncertainty: 0.9 · selection: Overall rows; not per-persona averages. COMP is a separate metric and does not determine this ranking. · tasks: 100

π-Bench research team ↗

MiniMax M2.755.6%

Overall · PROC

completeness score: 60 · completeness uncertainty: 1.8 · metric: Avg PROC · personas: 5 · proc uncertainty: 3.2 · selection: Overall rows; not per-persona averages. COMP is a separate metric and does not determine this ranking. · tasks: 100

π-Bench research team ↗

DeepSeek V3.253.3%

Overall · PROC

completeness score: 57.8 · completeness uncertainty: 3 · metric: Avg PROC · personas: 5 · proc uncertainty: 1.9 · selection: Overall rows; not per-persona averages. COMP is a separate metric and does not determine this ranking. · tasks: 100

π-Bench research team ↗

Other benchmarks

Humanity’s Last Exam

Humanity’s Last Exam evaluates advanced knowledge and reasoning across a broad range of academic subjects using expert-written questions.

BrowseComp

BrowseComp evaluates web research agents on questions with hard-to-find but verifiable answers.

Gaia2

Gaia2 evaluates assistants coordinating tasks across simulated email, calendars, shopping, contacts, and other apps.