AssistantBench
AssistantBench evaluates AI agents on realistic, time-consuming, and automatically verifiable tasks. It consists of 214 tasks that are based on real human needs and require several minutes of human browsing.
Paper: AssistantBench: Can Web Agents Solve Realistic and Time-Consuming Tasks? (Yoran et al., 2024)
Key Features of AssistantBench
Realistic, Time-Consuming Web Tasks
Tasks are based on real human needs and require multiple steps (several minutes) of web browsing to solve (e.g., finding specific business information, analyzing market trends, planning travel, etc.).
Diverse Domains
Tasks were created by everyday users, crowdworkers, and domain experts and cover a wide variety of domains.
Automatic Evaluation
Tasks were designed with closed-form, verifiable answers, allowing for automatic evaluations.
AssistantBench Leaderboard
| Rank | Agent | Models |
Verified
Verified Results
Results have been reproduced by the HAL team |
Accuracy
Accuracy
Confidence intervals show the min-max values across runs for those agents where multiple runs are available |
Cost (USD)
Total Cost
Total API cost for running the agent on all tasks. Confidence intervals show the min-max values across runs for those agents where multiple runs are available |
Runs
Number of Runs
The number of runs for this agent submitted to the leaderboard. To submit multiple evaluations, rerun the same agent and set the same agent name |
Traces |
|---|---|---|---|---|---|---|---|
| 1 |
Browser-Use
|
o3 Medium (April 2025) | ✓ | 42.83% (-4.01/+4.01) | $13.60 (-1.55/+1.55) | 2 | Download |
| 2 |
HAL Generalist Agent
|
DeepSeek V3 | ✓ | 37.50% | $0.92 | 1 | Download |
| 3 |
Browser-Use
|
GPT-5 Medium (August 2025) | ✓ | 35.88% (-0.65/+0.65) | $41.76 (-0.07/+0.07) | 2 | Download |
| 4 |
HAL Generalist Agent
|
o3 Medium (April 2025) | ✓ | 31.07% | $6.19 | 1 | Download |
| 5 |
HAL Generalist Agent
|
GPT-4.1 (April 2025) | ✓ | 28.95% | $10.83 | 1 | Download |
| 6 |
HAL Generalist Agent
|
Claude-3.7 Sonnet High (February 2025) | ✓ | 28.75% | $10.56 | 1 | Download |
| 7 |
Browser-Use
|
o4-mini Low (April 2025) | ✓ | 28.05% | $9.22 | 1 | Download |
| 8 |
HAL Generalist Agent
|
Claude Opus 4.1 (August 2025) | ✓ | 26.55% | $108.47 | 1 | Download |
| 9 |
HAL Generalist Agent
|
o4-mini High (April 2025) | ✓ | 24.83% | $8.84 | 1 | Download |
| 10 |
Browser-Use
|
o4-mini High (April 2025) | ✓ | 23.84% | $16.39 | 1 | Download |
| 11 |
HAL Generalist Agent
|
Claude-3.7 Sonnet (February 2025) | ✓ | 17.99% | $9.71 | 1 | Download |
| 12 |
HAL Generalist Agent
|
DeepSeek R1 | ✓ | 17.91% | $5.14 | 1 | Download |
| 13 |
Browser-Use
|
GPT-4.1 (April 2025) | ✓ | 17.39% | $14.15 | 1 | Download |
| 14 |
Browser-Use
|
Claude-3.7 Sonnet (February 2025) | ✓ | 16.69% | $56.00 | 1 | Download |
| 15 |
HAL Generalist Agent
|
o4-mini Low (April 2025) | ✓ | 15.98% | $2.11 | 1 | Download |
| 16 |
HAL Generalist Agent
|
Claude Opus 4.1 High (August 2025) | ✓ | 14.91% | $112.61 | 1 | Download |
| 17 |
HAL Generalist Agent
|
Gemini 2.0 Flash | ✓ | 14.40% | $1.40 | 1 | Download |
| 18 |
Browser-Use
|
Claude-3.7 Sonnet High (February 2025) | ✓ | 13.08% | $16.13 | 1 | Download |
| 19 |
Browser-Use
|
Claude Opus 4.1 (August 2025) | ✓ | 7.26% | $385.43 | 1 | Download |
| 20 |
Browser-Use
|
DeepSeek V3 | ✓ | 2.03% | $2.94 | 1 | Download |
| 21 |
Browser-Use
|
DeepSeek R1 | ✓ | 0.00% | $1.36 | 1 | Download |
Accuracy vs. Cost Frontier for AssistantBench
This plot shows the relationship between an agent's performance and its token cost. The Pareto frontier (dashed line) represents the current state-of-the-art trade-off. The error bars indicate min-max values across runs.
Heatmap for AssistantBench
The heatmap visualizes success rates across tasks and agents. Colorscale shows the fraction of times a task was solved across reruns of the same agent. The "any agent" performance indicates the level of saturation of the benchmark and gives a sense of overall progress.
Total Completion Tokens Used per Agent
The bar chart shows the total completion tokens used by each agent, with the height of each bar representing the total number of completion tokens used across all tasks. Secondary models usually contribute a relatively small amount of tokens in comparison, and are used for RAG or image processing only.
Token Pricing Configuration
Adjust token prices to see how they affect the total cost calculations in the leaderboard and plots.
Additional Resources
Getting Started
Want to evaluate your agent on AssistantBench ? Follow our comprehensive guide to get started:
View Documentation