The quality of AI products and services depends intricately on the capabilities of the LLMs that power them, and as these models continue to improve, we have to regularly revise our choices.
To keep you up to date on how new models are performing and which ones lead each category, we've put together a neat list of benchmarks and datasets.
Leaderboards
LM Arena
Subjects:
Chat models, Reasoning models
Topics:
reasoning
language understanding
user preference
LM Arena is an open platform created by researchers from UC Berkeley that lets users pose questions and compare responses from two different Large Language Models without knowing which model produced which answer. The aggregate results of these head-to-head "battles" generate a ranking and leaderboard, similar to the ELO rating system used in chess.
Scale Showdown
Subjects:
Chat models, Reasoning models
Topics:
user preference
language
demographics
real-world use
Scale Showdown ranks models using blind, optional comparisons gathered during real conversations. Its results can be explored by dimensions such as language, age, continent, education, profession, and prompt type, providing a broader view of user preference across populations.
Artificial Analysis
Subjects:
Chat models, Reasoning models, Code models
Topics:
intelligence
coding
speed
latency
price
Artificial Analysis independently evaluates models and combines capability results with practical measurements such as output speed, latency, price, and token usage. Its standardized runs make it useful for comparing both model quality and the cost of deploying it.
LiveBench
Subjects:
Chat models, Reasoning models, Code models
Topics:
math
coding
reasoning
language
instruction following
data analysis
LiveBench is an evolving benchmark for Large Language Models that aims to avoid “training set leakage” (contamination) by continuously introducing fresh, recently sourced tasks. Each question in LiveBench is paired with a verifiable ground-truth answer, so that model outputs can be scored automatically — no external judge (human or AI) is needed.
HELM
Subjects:
Chat models, Reasoning models, Multimodal models
Topics:
capabilities
safety
robustness
fairness
efficiency
Stanford's Holistic Evaluation of Language Models provides transparent, reproducible evaluations across capabilities, safety, multimodality, and specialist domains. The project entered maintenance mode in June 2026, so its established methodology and archived results are more useful than it is for comparing the newest models.
Humanity's Last Exam
Subjects:
Reasoning models, Multimodal models
Topics:
expert knowledge
reasoning
calibration
multimodality
Humanity's Last Exam measures expert-level knowledge and reasoning using difficult questions from mathematics, the sciences, humanities, and other disciplines. Its leaderboard reports both accuracy and calibration, including a multimodal subset with diagrams and figures.
MMLU-Pro
Subjects:
Chat models, Reasoning models
Topics:
knowledge
reasoning
science
humanities
MMLU-Pro is a more difficult and robust successor to MMLU. It expands the number of answer choices, emphasizes reasoning, and covers a broad range of academic disciplines while reducing sensitivity to prompt variations and chance performance.
LongBench v2
Subjects:
Chat models, Reasoning models, Long-context models
Topics:
long context
document understanding
reasoning
retrieval
LongBench v2 evaluates deep understanding and reasoning over realistic contexts ranging from thousands to millions of words. Its tasks cover documents, dialogue histories, code repositories, in-context learning, and structured data rather than simple information retrieval alone.
MMMU-Pro
Subjects:
Multimodal models, Reasoning models
Topics:
multimodality
visual understanding
expert knowledge
reasoning
MMMU-Pro tests expert-level reasoning across interleaved text and images such as charts, diagrams, maps, tables, and technical figures. It strengthens the original MMMU benchmark against shortcuts that let models answer without genuinely understanding the visual information.
Berkeley Function-Calling Leaderboard
Subjects:
Chat models, AI agents
Topics:
tool use
function calling
agent memory
web search
The Berkeley Function-Calling Leaderboard evaluates whether models can select and invoke tools accurately. Its tests span single- and multi-turn calls, relevance detection, web search, memory management, format sensitivity, latency, and cost.
τ-bench
Subjects:
AI agents, Voice agents
Topics:
tool use
conversation
policy following
knowledge retrieval
τ-bench evaluates agents that must converse with users, follow policies, retrieve knowledge, and call tools in realistic enterprise workflows. Its evolving family includes multi-party interaction, knowledge-heavy tasks, and real-time voice conversations.
SWE-bench
Subjects:
Code models, Coding agents
Topics:
bug fixing
repository comprehension
multilingual code
multimodality
SWE-bench evaluates systems on real GitHub issues that require changes to existing repositories. Its official leaderboards include Verified, Multilingual, Multimodal, Lite, and Full tracks, making results across different software-engineering settings easier to compare.
LiveSWEBench
Subjects:
Code models, Coding agents
Topics:
code editing
feature implementation
bug fixing
repository comprehension
LiveSWEBench is a live, continuously updated benchmark that evaluates the software engineering (SWE) capabilities of AI agents in real-world conditions. It draws tasks from active repositories and scores agents on their ability to propose correct, executable changes.
Terminal-Bench
Subjects:
Coding agents, Computer-use agents
Topics:
terminal use
system administration
coding
end-to-end tasks
Terminal-Bench evaluates agents on complete tasks inside real terminal environments. It covers software engineering, system administration, data processing, model training, and security work that requires more than generating an isolated answer or patch.