OPT-BENCH
Agents iteratively improve solutions to 20 machine-learning tasks and 10 combinatorial optimization problems using execution feedback. Win counts, invalid-solution rates, ranks, and improvement rates measure optimization performance.
Results depend on the supplied OPT-Agent refinement protocol and the number of iterations.
- Made by
- Tongji University, Shanghai AI Laboratory, Nanjing University and others
- Released
- June 2025
- Kind
- Scientific code
- Level
- University
- Scored
- Checked automatically
- Answers
- Code
- Tasks
- All public
- Measures
- Capability
- Language
- English
- Interaction
- Agentic
- Environment
- Code sandbox
- Refresh
- One fixed set
- Run by
- Anyone
- In views
- Agents
- Strongest model tested
-
Qwen3-235B-A22B-Thinking Jul 2025 #108 on the capability index
- Last model added
- August 2025
- Scores refreshed
- November 20th, 2026
In the authors’ words
Large Language Models (LLMs) have shown remarkable capabilities in solving diverse tasks. However, their proficiency in iteratively optimizing complex solutions through learning from previous feedback remains insufficiently explored. To bridge this gap, we present OPT-BENCH, a comprehensive benchmark designed to evaluate LLM agents on large-scale search space optimization problems. OPT-BENCH includes 20 real-world machine learning tasks sourced from Kaggle and 10 classical NP problems, offering a diverse and challenging environment for assessing LLM agents on iterative reasoning and solution refinement. To enable rigorous evaluation, we introduce OPT-Agent, an end-to-end optimization framework that emulates human reasoning when tackling complex problems by generating, validating, and iteratively improving solutions through leveraging historical feedback. Through extensive experiments on 9 state-of-the-art LLMs from 6 model families, we analyze the effects of optimization iterations, temperature settings, and model architectures on solution quality and convergence. Our results demonstrate that incorporating historical context significantly enhances optimization performance across both ML and NP tasks. All datasets, code, and evaluation tools are open-sourced to promote further research in advancing LLM-driven optimization and iterative reasoning. Project page: https://github.com/OliverLeeXZ/OPT-BENCH.
DeepSeek-V3.1 leads with 0.69 among models released by August 2025 DeepSeek-V3.1 leads with 0.69 among models released by August 2025
Best score per entry from the benchmark's paper, read on November 20th, 2026. A paper’s table doesn’t add newer models. The paper reports no overall score, so the headline is the plain mean of Expert Gap on the ML and NP tracks. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders. This view shows only models we track, each at its best.
The newest model with a score here came out in August 2025, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:
- GPT-6.2 Astra
- Claude Fable 5.2
-
Gemini 4 Carbon
- GPT-6 Astra
- and 16 more
- 0.69
DeepSeek-V3.1
- 0.69 o3-mini
- 0.69
Qwen3-235B-A22B-Thinking
- 0.63 Grok 3
- 0.61
Qwen3-30B-A3B
- 0.59 Claude 3.7 Sonnet
- 0.59
Qwen3-32B
- 0.58
Qwen3-235B-A22B-Instruct
- 0.56 GPT-4.1
- 0.54 GPT-4o
- 0.51
Qwen3-8B
- 0.50 Claude 3.5 Sonnet
- 0.49
Gemini 2.0 Flash
- 0.48 Qwen2.5-32B-Instruct
- 0.46
Qwen2.5-72B
- 0.39 Qwen2.5-14B-Instruct
- 0.28 Qwen2.5-7B-Instruct
- 0.27
InternLM3-8B-Instruct
- 0.18 Qwen2.5-3B-Instruct
- 1
DeepSeek-V3.1 20 steps 0.69
- 2 o3-mini 20 steps 0.69
- 3
Qwen3-235B-A22B-Thinking Jul 2025, 20 steps 0.69
- 4 Grok 3 20 steps 0.63
- 5
Qwen3-30B-A3B 20 steps 0.61
- 6 Claude 3.7 Sonnet 20 steps 0.59
- 7
Qwen3-32B 20 steps 0.59
- 8
Qwen3-235B-A22B-Instruct Jul 2025, 20 steps 0.58
- 9 GPT-4.1 20 steps 0.56
- 10 GPT-4o Aug 2024, 20 steps 0.54
- 11
Qwen3-8B 20 steps 0.51
- 12 Claude 3.5 Sonnet 20 steps 0.50
- 13
Gemini 2.0 Flash Feb 2025, 20 steps 0.49
- 14 Qwen2.5-32B-Instruct 20 steps 0.48
- 15
Qwen2.5-72B 20 steps 0.46
- 16 Qwen2.5-14B-Instruct 20 steps 0.39
- 17 Qwen2.5-7B-Instruct 20 steps 0.28
- 18
InternLM3-8B-Instruct 20 steps 0.27
- 19 Qwen2.5-3B-Instruct 20 steps 0.18
Source: The OPT-BENCH paper
- 1
DeepSeek-V3.1 20 steps 0.69
- 2 o3-mini 20 steps 0.69
- 3
Qwen3-235B-A22B-Thinking Jul 2025, 20 steps 0.69
- 4 Grok 3 20 steps 0.63
- 5
Qwen3-30B-A3B 20 steps 0.61
- 6 Claude 3.7 Sonnet 20 steps 0.59
- 7
Qwen3-32B 20 steps 0.59
- 8
Qwen3-235B-A22B-Instruct Jul 2025, 20 steps 0.58
- 9 GPT-4.1 20 steps 0.56
- 10 GPT-4o Aug 2024, 20 steps 0.54
- 11
Qwen3-8B 20 steps 0.51
- 12 Claude 3.5 Sonnet 20 steps 0.50
- 13
Gemini 2.0 Flash Feb 2025, 20 steps 0.49
- 14
Qwen2.5-72B 20 steps 0.46
- 15
InternLM3-8B-Instruct 20 steps 0.27
Source: The OPT-BENCH paper
| # | Entry | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | | 0.69 | 20 steps, OPT-Agent | Aug 2025 | Official leaderboard |
| 2 | o3-mini | 0.69 | 20 steps, OPT-Agent | Jan 2025 | Official leaderboard |
| 3 | | 0.69 | 20 steps, OPT-Agent | Jul 2025 | Official leaderboard |
| 4 | Grok 3 | 0.63 | 20 steps, OPT-Agent | Apr 2025 | Official leaderboard |
| 5 | | 0.61 | 20 steps, OPT-Agent | Apr 2025 | Official leaderboard |
| 6 | Claude 3.7 Sonnet | 0.59 | 20 steps, OPT-Agent | Feb 2025 | Official leaderboard |
| 7 | | 0.59 | 20 steps, OPT-Agent | Apr 2025 | Official leaderboard |
| 8 | | 0.58 | 20 steps, OPT-Agent | Jul 2025 | Official leaderboard |
| 9 | GPT-4.1 | 0.56 | 20 steps, OPT-Agent | Apr 2025 | Official leaderboard |
| 10 | GPT-4o Aug 2024 | 0.54 | 20 steps, OPT-Agent | Aug 2024 | Official leaderboard |
| 11 | | 0.51 | 20 steps, OPT-Agent | Apr 2025 | Official leaderboard |
| 12 | Claude 3.5 Sonnet | 0.50 | 20 steps, OPT-Agent | Jun 2024 | Official leaderboard |
| 13 | | 0.49 | 20 steps, OPT-Agent | Feb 2025 | Official leaderboard |
| 14 | Qwen2.5-32B-Instruct | 0.48 | 20 steps, OPT-Agent | Official leaderboard | |
| 15 | | 0.46 | 20 steps, OPT-Agent | Sep 2024 | Official leaderboard |
| 16 | Qwen2.5-14B-Instruct | 0.39 | 20 steps, OPT-Agent | Official leaderboard | |
| 17 | Qwen2.5-7B-Instruct | 0.28 | 20 steps, OPT-Agent | Official leaderboard | |
| 18 | | 0.27 | 20 steps, OPT-Agent | Jan 2025 | Official leaderboard |
| 19 | Qwen2.5-3B-Instruct | 0.18 | 20 steps, OPT-Agent | Official leaderboard |
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | | 0.69 | 20 steps, OPT-Agent | Aug 2025 | Official leaderboard |
| 2 | o3-mini | 0.69 | 20 steps, OPT-Agent | Jan 2025 | Official leaderboard |
| 3 | | 0.69 | 20 steps, OPT-Agent | Jul 2025 | Official leaderboard |
| 4 | Grok 3 | 0.63 | 20 steps, OPT-Agent | Apr 2025 | Official leaderboard |
| 5 | | 0.61 | 20 steps, OPT-Agent | Apr 2025 | Official leaderboard |
| 6 | Claude 3.7 Sonnet | 0.59 | 20 steps, OPT-Agent | Feb 2025 | Official leaderboard |
| 7 | | 0.59 | 20 steps, OPT-Agent | Apr 2025 | Official leaderboard |
| 8 | | 0.58 | 20 steps, OPT-Agent | Jul 2025 | Official leaderboard |
| 9 | GPT-4.1 | 0.56 | 20 steps, OPT-Agent | Apr 2025 | Official leaderboard |
| 10 | GPT-4o Aug 2024 | 0.54 | 20 steps, OPT-Agent | Aug 2024 | Official leaderboard |
| 11 | | 0.51 | 20 steps, OPT-Agent | Apr 2025 | Official leaderboard |
| 12 | Claude 3.5 Sonnet | 0.50 | 20 steps, OPT-Agent | Jun 2024 | Official leaderboard |
| 13 | | 0.49 | 20 steps, OPT-Agent | Feb 2025 | Official leaderboard |
| 14 | | 0.46 | 20 steps, OPT-Agent | Sep 2024 | Official leaderboard |
| 15 | | 0.27 | 20 steps, OPT-Agent | Jan 2025 | Official leaderboard |
Rankings by part
We read 2 parts from the benchmark’s paper: OPT-BENCH-ML Expert Gap and OPT-BENCH-NP Expert Gap. Each ranks models on one part of the test. o3-mini leads OPT-BENCH-ML Expert Gap and DeepSeek-V3.1 leads OPT-BENCH-NP Expert Gap. Arrows show how far a model moved from its overall rank among the models we track.
OPT-BENCH-ML Expert Gap
15 models we track have a score here.
- 1 o3-mini ↑1 0.65
- 2 Grok 3 ↑2 0.63
- 3
Qwen3-235B-A22B-Thinking Jul 2025 0.62
- 4 GPT-4o Aug 2024 ↑6 0.61
- 5
Qwen3-235B-A22B-Instruct Jul 2025 ↑3 0.60
- 6
DeepSeek-V3.1 ↓5 0.60
- 7 GPT-4.1 ↑2 0.58
- 8
Qwen3-32B ↓1 0.58
- 9
Qwen3-30B-A3B ↓4 0.54
- 10
Gemini 2.0 Flash Feb 2025 ↑3 0.53
OPT-BENCH-NP Expert Gap
15 models we track have a score here.
- 1
DeepSeek-V3.1 0.79
- 2
Qwen3-235B-A22B-Thinking Jul 2025 ↑1 0.76
- 3 o3-mini ↓1 0.73
- 4
Qwen3-30B-A3B ↑1 0.69
- 5 Claude 3.7 Sonnet ↑1 0.66
- 6
Qwen3-8B ↑5 0.63
- 7 Grok 3 ↓3 0.62
- 8
Qwen3-32B ↓1 0.60
- 9
Qwen3-235B-A22B-Instruct Jul 2025 ↓1 0.57
- 10 GPT-4.1 ↓1 0.53
Scores at each effort level
Models the leaderboard ran at two or more effort levels. We don’t have enough cost data here to say what the extra effort costs.
Source: The OPT-BENCH paper
| Model | Effort | Harness | Score | Cost per task | Output tokens |
|---|---|---|---|---|---|
| 20 steps | OPT-Agent | 0.69 | |||
| o3-mini | 20 steps | OPT-Agent | 0.69 | ||
| 20 steps | OPT-Agent | 0.69 | |||
| 10 steps | OPT-Agent | 0.68 | |||
| o3-mini | 10 steps | OPT-Agent | 0.68 | ||
| 10 steps | OPT-Agent | 0.65 | |||
| o3-mini | 5 steps | OPT-Agent | 0.64 | ||
| 5 steps | OPT-Agent | 0.63 | |||
| 5 steps | OPT-Agent | 0.63 | |||
| Grok 3 | 20 steps | OPT-Agent | 0.63 | ||
| 20 steps | OPT-Agent | 0.61 | |||
| Claude 3.7 Sonnet | 20 steps | OPT-Agent | 0.59 | ||
| 10 steps | OPT-Agent | 0.59 | |||
| 20 steps | OPT-Agent | 0.59 | |||
| 20 steps | OPT-Agent | 0.58 | |||
| Claude 3.7 Sonnet | 10 steps | OPT-Agent | 0.56 | ||
| Grok 3 | 10 steps | OPT-Agent | 0.56 | ||
| GPT-4.1 | 20 steps | OPT-Agent | 0.56 | ||
| GPT-4o Aug 2024 | 20 steps | OPT-Agent | 0.54 | ||
| 10 steps | OPT-Agent | 0.54 | |||
| Claude 3.7 Sonnet | 5 steps | OPT-Agent | 0.53 | ||
| 5 steps | OPT-Agent | 0.53 | |||
| 10 steps | OPT-Agent | 0.53 | |||
| GPT-4o Aug 2024 | 10 steps | OPT-Agent | 0.51 | ||
| Grok 3 | 5 steps | OPT-Agent | 0.51 | ||
| 20 steps | OPT-Agent | 0.51 | |||
| Claude 3.5 Sonnet | 20 steps | OPT-Agent | 0.50 | ||
| 20 steps | OPT-Agent | 0.49 | |||
| GPT-4.1 | 10 steps | OPT-Agent | 0.48 | ||
| 5 steps | OPT-Agent | 0.48 | |||
| GPT-4o Aug 2024 | 5 steps | OPT-Agent | 0.47 | ||
| 20 steps | OPT-Agent | 0.46 | |||
| 10 steps | OPT-Agent | 0.46 | |||
| Claude 3.5 Sonnet | 10 steps | OPT-Agent | 0.46 | ||
| 10 steps | OPT-Agent | 0.45 | |||
| 5 steps | OPT-Agent | 0.42 | |||
| GPT-4.1 | 5 steps | OPT-Agent | 0.42 | ||
| 10 steps | OPT-Agent | 0.41 | |||
| Claude 3.5 Sonnet | 5 steps | OPT-Agent | 0.40 | ||
| 5 steps | OPT-Agent | 0.40 | |||
| 5 steps | OPT-Agent | 0.36 | |||
| 5 steps | OPT-Agent | 0.35 | |||
| 20 steps | OPT-Agent | 0.27 | |||
| 10 steps | OPT-Agent | 0.21 | |||
| 5 steps | OPT-Agent | 0.09 |
Scores at each effort level
Models the leaderboard ran at two or more effort levels. We don’t have enough cost data here to say what the extra effort costs.
Source: The OPT-BENCH paper
| Model | Effort | Harness | Score | Cost per task | Output tokens |
|---|---|---|---|---|---|
| 20 steps | OPT-Agent | 0.69 | |||
| o3-mini | 20 steps | OPT-Agent | 0.69 | ||
| 20 steps | OPT-Agent | 0.69 | |||
| 10 steps | OPT-Agent | 0.68 | |||
| o3-mini | 10 steps | OPT-Agent | 0.68 | ||
| 10 steps | OPT-Agent | 0.65 | |||
| o3-mini | 5 steps | OPT-Agent | 0.64 | ||
| 5 steps | OPT-Agent | 0.63 | |||
| 5 steps | OPT-Agent | 0.63 | |||
| Grok 3 | 20 steps | OPT-Agent | 0.63 | ||
| 20 steps | OPT-Agent | 0.61 | |||
| Claude 3.7 Sonnet | 20 steps | OPT-Agent | 0.59 | ||
| 10 steps | OPT-Agent | 0.59 | |||
| 20 steps | OPT-Agent | 0.59 | |||
| 20 steps | OPT-Agent | 0.58 | |||
| Claude 3.7 Sonnet | 10 steps | OPT-Agent | 0.56 | ||
| Grok 3 | 10 steps | OPT-Agent | 0.56 | ||
| GPT-4.1 | 20 steps | OPT-Agent | 0.56 | ||
| GPT-4o Aug 2024 | 20 steps | OPT-Agent | 0.54 | ||
| 10 steps | OPT-Agent | 0.54 | |||
| Claude 3.7 Sonnet | 5 steps | OPT-Agent | 0.53 | ||
| 5 steps | OPT-Agent | 0.53 | |||
| 10 steps | OPT-Agent | 0.53 | |||
| GPT-4o Aug 2024 | 10 steps | OPT-Agent | 0.51 | ||
| Grok 3 | 5 steps | OPT-Agent | 0.51 | ||
| 20 steps | OPT-Agent | 0.51 | |||
| Claude 3.5 Sonnet | 20 steps | OPT-Agent | 0.50 | ||
| 20 steps | OPT-Agent | 0.49 | |||
| GPT-4.1 | 10 steps | OPT-Agent | 0.48 | ||
| 5 steps | OPT-Agent | 0.48 | |||
| GPT-4o Aug 2024 | 5 steps | OPT-Agent | 0.47 | ||
| 20 steps | OPT-Agent | 0.46 | |||
| 10 steps | OPT-Agent | 0.46 | |||
| Claude 3.5 Sonnet | 10 steps | OPT-Agent | 0.46 | ||
| 10 steps | OPT-Agent | 0.45 | |||
| 5 steps | OPT-Agent | 0.42 | |||
| GPT-4.1 | 5 steps | OPT-Agent | 0.42 | ||
| 10 steps | OPT-Agent | 0.41 | |||
| Claude 3.5 Sonnet | 5 steps | OPT-Agent | 0.40 | ||
| 5 steps | OPT-Agent | 0.40 | |||
| 5 steps | OPT-Agent | 0.36 | |||
| 5 steps | OPT-Agent | 0.35 | |||
| 20 steps | OPT-Agent | 0.27 | |||
| 10 steps | OPT-Agent | 0.21 | |||
| 5 steps | OPT-Agent | 0.09 |
2 models have held number one in the benchmark’s paper since its release in June 2025
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: The OPT-BENCH paper
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official OPT-BENCH leaderboard
No lab model card reports it yet
We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.
More in Coding
- SWE-bench Verified In use 500 human-validated GitHub issues from 12 Python repos. Agent writes a patch that must pass hidden tests. Claude Opus 4.7 83.5%
- LiveCodeBench v6 In use Competitive programming problems collected after model cutoffs from LeetCode, AtCoder, Codeforces. Scored pass@1 on hidden tests.
DeepSeek-V4-Pro 93.5%
- SWE-bench Multilingual In use 300 GitHub-issue tasks across nine programming languages in SWE-bench format. Patch graded by repository tests. Claude Fable 5.2 96.6%
- DeepSWE In use 113 original long-horizon engineering tasks across 91 repos and five languages, written from scratch. Hand-written verifiers accept alternative solutions. GPT-6 Astra 74.1%
- SciCode In use Scientist-curated research coding problems split into subproblems across natural sciences. Code checked against test cases.
Seed2.1 Pro 59.8%
- Codeforces rating In use Model solves recent Codeforces contest problems. An Elo-style rating is computed from simulated contest performance.
DeepSeek V4.1 Flash 3,471
Data audit
8 of 9 checks pass, 1 warning. Look first at the Newest score check. The newest came 15 months ago, in August 2025, for DeepSeek-V3.1. Every benchmark
Catalog record 2 pass
- Pass Task access All tasks are public.
- Pass Confidence Checked against a primary source.
Scores 1 warning 2 pass 2 notes
- Pass Independent runs Scores from the benchmark’s paper (19 entries).
- Pass Models scored 15 models we track, plus 4 other entries on its leaderboard.
- Warning Newest score The newest came 15 months ago, in August 2025, for DeepSeek-V3.1.
- Note Frontier models None of the top 10 has a score. The strongest tested is Qwen3-235B-A22B-Thinking (#108).
- Note Headroom Scored as a plain number, with no fixed ceiling.
Leaderboard import 3 pass
- Pass Importer We read the board on arxiv.org: 45 rows for models we track and 4 other entries at the last read.
- Pass Last read November 20th, 2026.
- Pass Read worked The last read worked.
Consistency 1 pass 1 note
- Note Lab claims No model has both a lab number and an independent run.
- Pass Known issues No published audit found problems.