CORE-Bench Hard
CORE-Bench evaluates the ability of agents to computationally reproduce the results of published scientific papers. In CORE-Bench Hard, the agent is only given the codebase of the paper and must install all libraries and dependencies, run the code, and read through the output and figures to answer questions about the paper. This level is most akin to fully reproducing a paper and is the most realistic and challenging level.
Key Features of CORE-Bench
Scientific Papers
Tasks are based on actual published scientific papers, requiring agents to understand and reproduce real research.
Comprehensive Evaluation
Tests multiple capabilities: code understanding, dependency management, and scientific result interpretation.
Three Difficulty Levels
Tasks range from interpreting results (Easy) to full paper reproduction (Hard), allowing granular capability assessment.
CORE-Bench Hard Leaderboard
Costs are currently calculated without accounting for caching benefits.
Update: Running Opus 4.5 with an updated scaffold that uses Claude Code drastically outperforms the CORE-Agent scaffold we used, especially after fixing a few grading errors via manual scoring. We have now declared CORE-Bench solved.
| Rank | Scaffold |
Primary Model
Primary Model
This is the primary model used by the agent. In some cases, an embedding model is used for RAG, or a secondary model like GPT-4o for image processing. Note: For non-OpenAI reasoning models, the reasoning token budget is set at 1,024 (low), 2,048 (medium), and 4,096 (high). |
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 |
CORE-Agent
Pareto optimal
|
Claude Opus 4.1 (August 2025) | ✓ | 51.11% | $412.42 | 1 | Download |
| 2 |
CORE-Agent
Pareto optimal
|
Claude Sonnet 4.5 High (September 2025) | ✓ | 44.44% | $92.34 | 1 | Download |
| 3 | Claude Opus 4.5 High (November 2025) | ✓ | 42.22% | $152.66 | 1 | Download | |
| 4 | Claude Opus 4.5 (November 2025) | ✓ | 42.22% | $168.99 | 1 | Download | |
| 5 | Claude Opus 4.1 High (August 2025) | ✓ | 42.22% | $509.95 | 1 | Download | |
| 6 | Gemini 3 Pro Preview High (November 2025) | ✓ | 40.00% | $86.60 | 1 | Download | |
| 7 | Claude-3.7 Sonnet High (February 2025) | ✓ | 37.78% | $66.15 | 1 | Download | |
| 8 | Claude Sonnet 4.5 (September 2025) | ✓ | 37.78% | $97.15 | 1 | Download | |
| 9 |
HAL Generalist Agent
Pareto optimal
|
o4-mini High (April 2025) | ✓ | 35.56% | $45.37 | 1 | Download |
| 10 | Claude-3.7 Sonnet (February 2025) | ✓ | 35.56% | $73.04 | 1 | Download | |
| 11 | Gemini 3 Pro Preview High (November 2025) | ✓ | 35.56% | $101.27 | 1 | Download | |
| 12 | Claude Opus 4.1 (August 2025) | ✓ | 35.56% | $375.11 | 1 | Download | |
| 13 | Claude Sonnet 4.5 (September 2025) | ✓ | 33.33% | $85.19 | 1 | Download | |
| 14 | Claude Sonnet 4 High (May 2025) | ✓ | 33.33% | $100.48 | 1 | Download | |
| 15 | GPT-4.1 (April 2025) | ✓ | 33.33% | $107.36 | 1 | Download | |
| 16 | Claude Opus 4.5 (November 2025) | ✓ | 33.33% | $127.41 | 1 | Download | |
| 17 | Claude Opus 4.1 High (August 2025) | ✓ | 33.33% | $358.47 | 1 | Download | |
| 18 | Claude-3.7 Sonnet (February 2025) | ✓ | 31.11% | $56.64 | 1 | Download | |
| 19 | Claude Opus 4.5 High (November 2025) | ✓ | 31.11% | $112.38 | 1 | Download | |
| 20 | Claude Sonnet 4 (May 2025) | ✓ | 28.89% | $50.27 | 1 | Download | |
| 21 | Claude Sonnet 4.5 High (September 2025) | ✓ | 28.89% | $87.77 | 1 | Download | |
| 22 | GPT-5 Medium (August 2025) | ✓ | 26.67% | $31.76 | 1 | Download | |
| 23 | o4-mini High (April 2025) | ✓ | 26.67% | $61.35 | 1 | Download | |
| 24 | Claude-3.7 Sonnet High (February 2025) | ✓ | 24.44% | $72.47 | 1 | Download | |
| 25 | o3 Medium (April 2025) | ✓ | 24.44% | $120.47 | 1 | Download | |
| 26 | GPT-4.1 (April 2025) | ✓ | 22.22% | $58.32 | 1 | Download | |
| 27 | o3 Medium (April 2025) | ✓ | 22.22% | $88.34 | 1 | Download | |
| 28 | Gemini 2.5 Pro Preview (March 2025) | ✓ | 22.22% | $182.34 | 1 | Download | |
| 29 |
CORE-Agent
Pareto optimal
|
DeepSeek V3.1 (August 2025) | ✓ | 20.00% | $12.55 | 1 | Download |
| 30 | DeepSeek V3 (March 2025) | ✓ | 17.78% | $25.26 | 1 | Download | |
| 31 | o4-mini Low (April 2025) | ✓ | 17.78% | $31.79 | 1 | Download | |
| 32 | o4-mini Low (April 2025) | ✓ | 15.56% | $22.50 | 1 | Download | |
| 33 | GPT-OSS-120B (August 2025) | ✓ | 11.11% | $4.21 | 1 | Download | |
| 34 | GPT-OSS-120B High (August 2025) | ✓ | 11.11% | $4.21 | 1 | Download | |
| 35 | Gemini 2.0 Flash (February 2025) | ✓ | 11.11% | $12.46 | 1 | Download | |
| 36 | GPT-5 Medium (August 2025) | ✓ | 11.11% | $29.75 | 1 | Download | |
| 37 | Claude Haiku 4.5 (October 2025) | ✓ | 11.11% | $43.93 | 1 | Download | |
| 38 |
HAL Generalist Agent
Pareto optimal
|
GPT-OSS-120B High (August 2025) | ✓ | 8.89% | $2.05 | 1 | Download |
| 39 | GPT-OSS-120B (August 2025) | ✓ | 8.89% | $2.79 | 1 | Download | |
| 40 | DeepSeek V3 (March 2025) | ✓ | 8.89% | $4.69 | 1 | Download | |
| 41 | DeepSeek R1 (May 2025) | ✓ | 8.89% | $7.77 | 1 | Download | |
| 42 | DeepSeek R1 (January 2025) | ✓ | 6.67% (-2.22/+2.22) | $81.11 (-46.45/+46.45) | 2 | Download | |
| 43 | DeepSeek R1 (January 2025) | ✓ | 4.45% (-2.22/+2.22) | $24.95 (-11.07/+22.15) | 2 | Download | |
| 44 | Gemini 2.0 Flash (February 2025) | ✓ | 4.44% | $7.06 | 1 | Download | |
| 45 | Gemini 2.5 Pro Preview (March 2025) | ✓ | 4.44% | $30.38 | 1 | Download |
Accuracy vs. Cost Frontier for CORE-Bench
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 CORE-Bench
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.
Model Performance Over Time
Track how model accuracy has evolved over time since their release dates. Each point represents the best performance achieved by that model on CORE-Bench Hard.
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 CORE-Bench? Follow our comprehensive guide to get started:
View Documentation