Benchmark leaderboards - updated October 17, 2026
AI benchmark leaderboards
Developers Digest tracks 16 AI benchmarks across 21 leaderboards with 488 sourced entries, each copied with the setting it was run under and a link to where it was read. 14 of the 16 come straight from an official leaderboard, and GPQA Diamond and SimpleQA include vendor-reported numbers, labelled on every row.
Each board keeps its own order and its own setting. Scores are never averaged across benchmarks or compared across boards.
All benchmarks
25 entries
Aider Polyglot
Aider's polyglot benchmark tests LLMs on 225 challenging Exercism coding exercises across C++, Go, Java, JavaScript, Python, and Rust.
- First on Aider Polyglot
- gpt-5 (high)88.0%
Verified Leaderboard27 entries
AIME
MathArena evaluates LLMs on the latest math competitions and olympiads; for AIME 2026 each model is run 4 times on each of the 30 final-answer problems and the average accuracy is reported.
- Tied first (2) on AIME 2026 (MathArena)
- GPT-5.5 (xhigh)100.00%
Verified Leaderboard50 entries - 2 boards
ARC-AGI
ARC-AGI-2 is designed to stress-test the capabilities of state-of-the-art AI reasoning systems, and ARC-AGI-3 is an interactive reasoning benchmark which challenges AI agents to explore novel environments, acquire goals on the fly, build adaptable world models, and learn continuously (arcprize.org).
- First on ARC-AGI-2
- GPT-6 Astra (Max)95.0%
- First on ARC-AGI-3
- GPT-6 Astra (High)99.9%
Verified Leaderboard25 entries
Artificial Analysis Intelligence Index
Artificial Analysis Intelligence Index combines a comprehensive suite of evaluation datasets to assess language model capabilities across reasoning, knowledge, maths and programming.
- First on Artificial Analysis Intelligence Index
- Claude Opus 5.5 (max with fallback)58
Verified Leaderboard43 entries - 2 boards
DeepSWE, NL2Repo and ProgramBench
DeepSWE measures frontier coding agents on original, long-horizon software engineering tasks, and ProgramBench gives agents only a compiled binary and its documentation and asks them to build a complete codebase that reproduces the program's behavior.
- First on DeepSWE v1.1
- gpt-6-astra74%±3%
- First on ProgramBench
- Claude Opus 5 (xhigh)4.5% resolved (Almost: 37.0%)
Verified Leaderboard20 entries
GPQA Diamond
GPQA is a dataset of 448 multiple-choice questions written by domain experts in biology, physics, and chemistry, and GPQA Diamond is its 198-question subset where both expert validators answered correctly and at most one of three non-experts did.
- Listed first on GPQA Diamond
- GPT-6 Astra96.0%
Verified Leaderboard25 entries
Humanity's Last Exam
Humanity's Last Exam is a multi-modal benchmark at the frontier of human knowledge, designed to be the final closed-ended academic benchmark of its kind with broad subject coverage.
- First on Humanity's Last Exam
- GPT 6 Astra54.80±1.94
Verified Leaderboard25 entries
LiveCodeBench
LiveCodeBench is a holistic and contamination-free evaluation benchmark of LLMs for code that continuously collects new problems over time.
- First on LiveCodeBench
- O4-Mini (High)80.2
Verified Leaderboard50 entries - 2 boards
LMArena
LMArena ranks AI models on two boards: Text, covering math, coding, creative writing and other open-ended tasks, and WebDev, covering front-end web development including multi-step agentic coding with tool use.
- First on LMArena Text (Text Arena Overall)
- Gemini 4 Argon (High) (Google)1524±9 (Preliminary)
- First on LMArena WebDev (Code Arena WebDev Overall)
- Claude Opus 5.5 (Max) (Anthropic)1807+14/-14
Verified Leaderboard25 entries
MLE-bench
The MLE-bench dataset is a collection of 75 Kaggle competitions which we use to evaluate the ML engineering capabilities of AI systems.
- First on MLE-bench
- Famou-Agent 2.0 (Gemini-3-Pro-Preview)64.44 ± 1.18
Verified Leaderboard26 entries
MMMU
MMMU evaluates multimodal models on massive multi-discipline tasks demanding college-level subject knowledge and deliberate reasoning, using 11.5K multimodal questions from college exams, quizzes, and textbooks across six disciplines, 30 subjects and 183 subfields.
- First on MMMU
- GPT-5.185.4*
Verified Leaderboard25 entries
OSWorld
OSWorld is a benchmark of 369 computer tasks involving real web and desktop apps in open domains, OS file I/O, and workflows spanning multiple applications, each with an initial state setup and a custom execution-based evaluation script, run in a real computer environment.
- First on OSWorld (OSWorld-Verified)
- claude-fable-5[1m] (Anthropic)86.0%
Verified Leaderboard25 entries
SimpleQA
SimpleQA is a benchmark that evaluates the ability of language models to answer short, fact-seeking questions.
- First on SimpleQA
- gpt-4.5-preview-2025-02-2762.5
Verified Leaderboard49 entries - 2 boards
SWE-bench (Verified and Pro)
SWE-bench Verified is a human-filtered subset of 500 real GitHub issues scored by % Resolved, and SWE-Bench Pro evaluates AI agents on realistic software engineering tasks scored by Resolve Rate, the percentage of tasks an agent resolves.
- Tied first (2) on SWE-bench Verified
- Claude 4.5 Opus79.20
- First on SWE-Bench Pro (Public Dataset)
- Muse Spark 1.161.50±3.10
Verified Leaderboard25 entries - 2 boards
tau-bench and tau2-bench
tau-bench measures how well AI agents converse with users, call tools, retrieve knowledge, and follow policy across enterprise customer-service domains, in text and voice.
- First on tau2-bench (Overall)
- Qwen3.5-397B-A17B (Alibaba Cloud)87.9%
- First on tau3-Banking (tau-knowledge)
- Qwen 3.8 Max (Qwen)55.2%
Verified Leaderboard23 entries
Terminal-Bench
Terminal-Bench is the benchmark for testing AI agents in real terminal environments, evaluating how well agents can handle real-world, end-to-end tasks (from compiling code to training models and setting up servers) autonomously.
- First on Terminal-Bench 4.0
- Opus 5.564.8% ± 3.1%
Verified Leaderboard
Method
How we source
Official leaderboards come first. Each entry is copied from the board as it was displayed on the date shown, with the harness, effort level or split it was run under. When no official board exists, the numbers come from the model maker's own model card or launch post and are labelled vendor-reported on every row. Third-party aggregators are not used, and caveats about each board, such as a board that has stopped updating, are shown on its page.
Sources read
- Aider Polyglotaider.chat -
- AIMEmatharena.ai -
- ARC-AGIarcprize.org -
- Artificial Analysis Intelligence Indexartificialanalysis.ai -
- DeepSWE, NL2Repo and ProgramBenchdeepswe.datacurve.ai, programbench.com -
- GPQA Diamondopenai.com, www-cdn.anthropic.com and 6 more -
- Humanity's Last Examlabs.scale.com -
- LiveCodeBenchlivecodebench.github.io -
- LMArenaarena.ai -
- MLE-benchgithub.com -
- MMMUmmmu-benchmark.github.io -
- OSWorldosworld-v1.xlang.ai -
- SimpleQAgithub.com -
- SWE-bench (Verified and Pro)swebench.com, labs.scale.com -
- tau-bench and tau2-benchtaubench.com -
- Terminal-Benchtbench.ai -
Newsletter
Get new benchmark results as they land
One email per week. New model releases, the leaderboard moves that matter and what they mean for what you build.