AssistantEval
Collection
AssistantEval by ORO. Published results and recorded trajectories. Website: assistanteval.com • 2 items • Updated
group stringclasses 2
values | rank int64 1 6 | id stringclasses 9
values | name stringclasses 9
values | provider stringclasses 5
values | model stringclasses 6
values | graded_runs int64 158 170 | tasks int64 34 34 | passed int64 81 133 | failed int64 27 76 | unsafe int64 1 10 | pass_rate float64 0.51 0.8 | pass_rate_95_low float64 0.44 0.73 | pass_rate_95_high float64 0.59 0.86 | checks_passed int64 1.52k 1.95k | checks_total int64 1.63k 2.06k | checks_share float64 0.92 0.98 | median_duration_s float64 19.7 275 | median_interaction_duration_s float64 13 103 | interaction_timed_runs int64 158 170 | consent_slip_free float64 0.94 1 | asking_slip_free float64 0.42 0.86 | grounded_slip_free float64 0.98 1 | truthful_slip_free float64 0.97 0.99 | reporting_slip_free float64 0.99 1 | done_slip_free float64 0.54 0.81 | consent_runs int64 139 164 | asking_runs int64 21 35 | grounded_runs int64 158 170 | truthful_runs int64 145 170 | reporting_runs int64 128 154 | done_runs int64 158 170 | consent_check_share float64 0.96 1 | asking_check_share float64 0.45 0.87 | grounded_check_share float64 0.99 1 | truthful_check_share float64 0.99 1 | reporting_check_share float64 1 1 | done_check_share float64 0.84 0.95 | consent_checks int64 183 226 | asking_checks int64 24 43 | grounded_checks int64 181 220 | truthful_checks int64 291 669 | reporting_checks int64 175 219 | done_checks int64 698 740 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
product | 1 | pally | Pally | null | null | 167 | 34 | 116 | 46 | 5 | 0.695 | 0.621 | 0.759 | 1,569 | 1,645 | 0.954 | 274.8 | 63.2 | 167 | 1 | 0.706 | 0.982 | 0.988 | 0.993 | 0.731 | 159 | 34 | 167 | 163 | 149 | 167 | 1 | 0.744 | 0.985 | 0.993 | 0.995 | 0.917 | 191 | 39 | 195 | 291 | 204 | 725 |
product | 2 | muse | Muse | null | null | 163 | 34 | 93 | 65 | 5 | 0.571 | 0.494 | 0.644 | 1,557 | 1,694 | 0.919 | 226 | 102.5 | 163 | 0.978 | 0.619 | 0.994 | 0.972 | 1 | 0.601 | 139 | 21 | 163 | 145 | 128 | 163 | 0.973 | 0.583 | 0.994 | 0.99 | 1 | 0.836 | 184 | 24 | 181 | 413 | 179 | 713 |
product | 3 | grok_bot | Grok Bot | null | null | 158 | 34 | 81 | 76 | 1 | 0.513 | 0.435 | 0.589 | 1,521 | 1,627 | 0.935 | 143.4 | 58 | 158 | 0.993 | 0.517 | 1 | 0.987 | 1 | 0.538 | 149 | 29 | 158 | 157 | 139 | 158 | 0.995 | 0.588 | 1 | 0.994 | 1 | 0.872 | 183 | 34 | 188 | 349 | 175 | 698 |
model | 1 | claude-opus-5.5 | Claude Opus 5.5 | Anthropic | anthropic/claude-opus-5.5 | 166 | 34 | 133 | 27 | 6 | 0.801 | 0.734 | 0.855 | 1,932 | 1,979 | 0.976 | 27.9 | 21.4 | 166 | 0.963 | 0.857 | 1 | 0.988 | 1 | 0.813 | 161 | 35 | 166 | 164 | 151 | 166 | 0.973 | 0.86 | 1 | 0.996 | 1 | 0.954 | 226 | 43 | 197 | 569 | 219 | 725 |
model | 2 | gpt-6-astra | GPT-6 Astra | OpenAI | openai/gpt-6-astra | 169 | 34 | 127 | 41 | 1 | 0.751 | 0.681 | 0.811 | 1,643 | 1,707 | 0.963 | 19.7 | 13 | 169 | 0.994 | 0.839 | 1 | 0.994 | 1 | 0.763 | 160 | 31 | 169 | 165 | 150 | 169 | 0.995 | 0.861 | 1 | 0.997 | 1 | 0.923 | 201 | 36 | 214 | 323 | 195 | 738 |
model | 3 | claude-fable-5.1 | Claude Fable 5.1 | Anthropic | anthropic/claude-fable-5.1 | 169 | 34 | 125 | 40 | 4 | 0.74 | 0.669 | 0.8 | 1,879 | 1,937 | 0.97 | 34.2 | 27 | 169 | 0.976 | 0.853 | 1 | 0.988 | 1 | 0.775 | 164 | 34 | 169 | 169 | 154 | 169 | 0.982 | 0.872 | 1 | 0.996 | 1 | 0.936 | 221 | 39 | 220 | 508 | 212 | 737 |
model | 4 | gemini-3.1-pro-preview | Gemini 3.1 Pro | Google | google/gemini-3.1-pro-preview | 165 | 34 | 109 | 50 | 6 | 0.661 | 0.585 | 0.728 | 1,656 | 1,739 | 0.952 | 22.9 | 17 | 165 | 0.975 | 0.419 | 1 | 0.994 | 1 | 0.679 | 161 | 31 | 165 | 165 | 149 | 165 | 0.98 | 0.5 | 1 | 0.997 | 1 | 0.917 | 203 | 36 | 198 | 382 | 196 | 724 |
model | 5 | grok-4.7 | Grok 4.7 | xAI | x-ai/grok-4.7 | 168 | 34 | 110 | 49 | 9 | 0.655 | 0.58 | 0.722 | 1,643 | 1,750 | 0.939 | 19.9 | 13 | 168 | 0.944 | 0.485 | 1 | 0.982 | 0.993 | 0.685 | 162 | 33 | 168 | 163 | 152 | 168 | 0.958 | 0.45 | 1 | 0.992 | 0.995 | 0.902 | 212 | 40 | 189 | 371 | 205 | 733 |
model | 6 | muse-spark-1.3 | Muse Spark 1.3 | Meta | meta/muse-spark-1.3 | 170 | 34 | 111 | 49 | 10 | 0.653 | 0.579 | 0.72 | 1,952 | 2,057 | 0.949 | 32.5 | 26 | 170 | 0.951 | 0.455 | 1 | 0.988 | 1 | 0.665 | 164 | 33 | 170 | 170 | 152 | 170 | 0.962 | 0.45 | 1 | 0.997 | 1 | 0.901 | 212 | 40 | 192 | 669 | 204 | 740 |
AssistantEval results
AssistantEval results are aggregate measurements of assistants completing synthetic, stateful tasks. This repository contains the published leaderboard CSV: one row per assistant, with task outcomes, behavior measures, grading counts, confidence intervals, and timing fields.
The CSV is the original public file from assistanteval.com/data/results.csv. The accompanying source index and methodology describe the published snapshot and grading process.
Snapshot
The source index was generated on October 6, 2026. It covers 9 assistants and 34 tasks. The CSV contains one aggregate row per assistant. It is a snapshot of completed evaluations; a later grading update does not imply that assistant conversations were rerun.
Data and use
Load data/results.csv with any CSV reader. The group field distinguishes products from models; graded_runs, tasks, passed, failed, and unsafe describe outcome counts. pass_rate is the share of runs passing every applicable check, and pass_rate_95_low / pass_rate_95_high give the published interval. Behavior fields report check shares and slip-free run shares; timing fields are in seconds. See the data guide for definitions.
The Hugging Face leaderboard configuration uses a test split to organize this table; it does not define a new benchmark split.
import pandas as pd
results = pd.read_csv("data/results.csv")
print(results[["name", "graded_runs", "pass_rate", "unsafe"]])
For a versioned download, cite the repository revision and the AssistantEval source. The corresponding public run records are documented separately at AssistantEval trajectories; they have different usage terms.
Method and limitations
Tasks run against synthetic services initialized to the same starting state. A simulated user supplies task-defined facts and approvals. Code checks actions, their order, values, and final state; recorded model judgments assess meaning, with unresolved judgments sent to human review. A run passes only when every applicable check passes. Paired tasks change a relevant fact or instruction, and repeated trials measure consistency.
These results describe the tested tasks, configurations, and snapshot. They should not be treated as a general ranking across all assistant use, or as evidence about model versions or settings beyond those documented by AssistantEval. See the full methodology.
License and citation
The leaderboard data is licensed under Creative Commons Attribution 4.0 International. Credit AssistantEval (assistanteval.com). See LICENSE.
Suggested citation:
AssistantEval. AssistantEval results (published October 6, 2026). https://assistanteval.com/data/results.csv. CC BY 4.0.
- Downloads last month
- 40