SWE-bench + • OproAI

SWE-bench is a benchmark for evaluating large language models on real world software issues collected from GitHub. Given a codebase and an issue, a language model is tasked with generating a patch that resolves the described problem. SWE-bench Verified is a human-validated subset that more reliably evaluates AI models’ ability to solve issues. International Olympiad in Informatics (IOI) competition features standardized and automated grading.

🏆 Leaderboard#

Model	SWE-bench	IOI	Organization	License
Claude Opus 5	97	91.7	Anthropic	Proprietary
DeepSeek-V4-Pro	96.4		DeepSeek	MIT
GPT-5.6 Sol	96.2	86.7	OpenAI	Proprietary
Grok 4.6	95.6		xAI	Proprietary
GLM-5.3	95.4		Z.ai	MIT
Claude Fable 5	95	72.3	Anthropic	Proprietary
Kimi K3	93.4		Moonshot	Kimi K3
DeepSeek-V4-Flash	88.8		DeepSeek	MIT
Claude Opus 4.8	88.6		Anthropic	Proprietary
Grok 4.5	86.6		xAI	Proprietary
Muse Spark 1.2	86.6	49.5	Meta	Proprietary
Qwen3.8-Max	85.6	73	Alibaba	Qwen3
GLM-5.2	82.8		Z.ai	MIT
GPT-5.5	82.6		OpenAI	Proprietary
Claude Opus 4.7	82	47.1	Anthropic	Proprietary
Muse Spark 1.1	82		Meta	Proprietary
Claude Sonnet 5	79.8		Anthropic	Proprietary
Gemini 3.6 Flash	79.8		Google	Proprietary
Gemini 3.1 Pro	79.8		Google	Proprietary
Qwen3.7-Max	79.4	46.8	Alibaba	Proprietary
GPT-5.4	79.2	67.8	OpenAI	Proprietary
Claude Opus 4.6	79.2	36.3	Anthropic	Proprietary
Kimi K2.7	78.2		Moonshot	Modified MIT
Gemini 3 Flash	76.2	39.1	Google	Proprietary
Grok 4.20	76.2	30.2	xAI	Proprietary
Claude Sonnet 4.6	76.2	21.5	Anthropic	Proprietary
GPT-5.3-Codex	75.2	54.8	OpenAI	Proprietary
Gemini 3 Pro	74.4	38.8	Google	Proprietary
GLM-5.1	74.4	32	Z.ai	MIT
Claude Opus 4.5	74.4	23.6	Anthropic	Proprietary
Kimi K2.6	74.2		Moonshot	Modified MIT
GPT-5.2	72.8	43.8	OpenAI	Proprietary
GLM-5	72.8	22	Z.ai	MIT
MiniMax M2.5	72.8	6.7	MiniMax	Apache 2.0
Kimi K2.5	70.8	17.7	Moonshot	Modified MIT
Grok 4	70.6	26.2	xAI	Proprietary
Claude Sonnet 4.5	70.6	18.5	Anthropic	Proprietary
GPT-5.1	70.5	21.5	OpenAI	Proprietary
GPT-5	70.1	20	OpenAI	Proprietary
Claude Opus 4.1	70.1	15.2	Anthropic	Proprietary
DeepSeek-V3.2	70	14.4	DeepSeek	MIT
Qwen3-Max	69.6	15.7	Alibaba	Proprietary
Claude Haiku 4.5	68.8	6.2	Anthropic	Proprietary
Claude Sonnet 4	68	6.5	Anthropic	Proprietary
Claude Opus 4	67.6		Anthropic	Proprietary
Qwen3-Coder-480B-A35B-Instruct	67		Alibaba	Apache 2.0
DeepSeek-V3.1	66		DeepSeek	MIT
Kimi K2	65.4	1.3	Moonshot	Modified MIT
GPT-5 (medium)	65		OpenAI	Proprietary
GLM-4.5	64.2		Z.ai	MIT
GPT-5 mini	59.8		OpenAI	Proprietary
o3	58.4		OpenAI	Proprietary
GLM-4.5-Air	57.6		Z.ai	MIT
Gemini 2.5 Pro	53.6	17.1	Google	Proprietary
Claude 3.7 Sonnet	52.8		Anthropic	Proprietary
Qwen3-Coder-30B-A3B-Instruct	51.6		Alibaba	Apache 2.0
GPT-4.1	48.6		OpenAI	Proprietary
o4-mini	45	5.3	OpenAI	Proprietary
DeepSeek-R1-0528	41.4		DeepSeek	MIT
DeepSeek-V3-0324	38.8	1.7	DeepSeek	MIT
GPT-5 nano	34.8		OpenAI	Proprietary
Gemini 2.5 Flash	28.7	3.9	Google	Proprietary
GPT-4.1 mini	23.9		OpenAI	Proprietary
GPT-4o	21.6		OpenAI	Proprietary
Llama 4 Maverick Instruct	21		Meta	Llama 4
Gemini 2.0 Flash	13.5		Google	Proprietary
Llama 4 Scout Instruct	9.1		Meta	Llama 4
Qwen2.5-Coder-32B-Instruct	9		Alibaba	Apache 2.0

SWE-bench Verified (100 turns)

IOI Benchmark

swe

👋 Overview#

SWE-bench tests AI systems’ ability to solve GitHub issues.

We collect 2,294 task instances by crawling Pull Requests and Issues from 12 popular Python repositories. Each instance is based on a pull request that (1) is associated with an issue, and (2) modified 1+ testing related files.

Per instance, we construct an execution environment (Docker Image) with the repository successfully installed at the commit that the Pull Request is based on. Without the Pull Request’s changes, a number of test(s) fail. After the Pull Request is merged, the same set of test(s) pass. These “Fail-to-Pass” tests are the primary signal for evaluation.

SWE-bench evaluation works as follows. Per task instance, an AI system is given the issue text. The AI system should then modify the codebase in order to resolve the described issues. When the AI system is finished, we run the aforementioned Fail-to-Pass tests to check if the issue was successfully resolved.

Code and data for the following works:

[ICLR 2025] SWE-bench Multimodal: Do AI Systems Generalize to Visual Software Domains?
[ICLR 2024 Oral] SWE-bench: Can Language Models Resolve Real-World GitHub Issues?

🚀 Set Up#

SWE-bench uses Docker for reproducible evaluations. Follow the instructions in the Docker setup guide to install Docker on your machine. If you’re setting up on Linux, we recommend seeing the post-installation steps as well.

Finally, to build SWE-bench from source, follow these steps:

git clone git@github.com:princeton-nlp/SWE-bench.git
cd SWE-bench
pip install -e .

bash

Test your installation by running:

python -m swebench.harness.run_evaluation \
    --predictions_path gold \
    --max_workers 1 \
    --instance_ids sympy__sympy-20590 \
    --run_id validate-gold

bash

ℹ️ Note

If using a MacOS M-series or other ARM-based systems, add --namespace '' to the above script. By default, the evaluation script pulls images (built for Linux) from DockerHub. Adding --namespace '' will cause evaluation images to be built locally instead.

💽 Usage#

Evaluate patch predictions on SWE-bench Lite with the following command:

python -m swebench.harness.run_evaluation \
    --dataset_name princeton-nlp/SWE-bench_Lite \
    --predictions_path <path_to_predictions> \
    --max_workers <num_workers> \
    --run_id <run_id>
    # use --predictions_path 'gold' to verify the gold patches
    # use --run_id to name the evaluation run
    # use --modal true to run on Modal

bash

This command will generate docker build logs (logs/build_images) and evaluation logs (logs/run_evaluation) in the current directory.

The final evaluation results will be stored in the evaluation_results directory.

⚠️ Warning

SWE-bench evaluation can be resource intensive We recommend running on an x86_64 machine with at least 120GB of free storage, 16GB of RAM, and 8 CPU cores. We recommend using fewer than min(0.75 * os.cpu_count(), 24) for --max_workers.

If running with Docker desktop, make sure to increase your virtual disk space to ~120 free GB. Set max_workers to be consistent with the above for the CPUs available to Docker.

Support for arm64 machines is experimental.

To see the full list of arguments for the evaluation harness, run:

python -m swebench.harness.run_evaluation --help

bash

✍️ Citation#

If you find our work helpful, please use the following citations.

@inproceedings{
    jimenez2024swebench,
    title={{SWE}-bench: Can Language Models Resolve Real-world Github Issues?},
    author={Carlos E Jimenez and John Yang and Alexander Wettig and Shunyu Yao and Kexin Pei and Ofir Press and Karthik R Narasimhan},
    booktitle={The Twelfth International Conference on Learning Representations},
    year={2024},
    url={https://openreview.net/forum?id=VTF8yNQM66}
}

@inproceedings{
    yang2024swebenchmultimodal,
    title={{SWE}-bench Multimodal: Do AI Systems Generalize to Visual Software Domains?},
    author={John Yang and Carlos E. Jimenez and Alex L. Zhang and Kilian Lieret and Joyce Yang and Xindi Wu and Ori Press and Niklas Muennighoff and Gabriel Synnaeve and Karthik R. Narasimhan and Diyi Yang and Sida I. Wang and Ofir Press},
    booktitle={The Thirteenth International Conference on Learning Representations},
    year={2025},
    url={https://openreview.net/forum?id=riTiq3i21b}
}

bibtex