These benchmarks compare general-purpose models on reasoning, facts, and instruction following. Use them to make a first shortlist. They do not show that a model works on your task.
Which benchmarks are useful for comparing general-purpose AI models?
Start with one open leaderboard or test harness. Then read the component benchmarks behind any composite score. Prefer new test sets that resist contamination. Check that the tested model version and settings match what you will deploy.
13
Benchmarks in this domain
Sources first · alphabetical collection order
Runnability describes access to a usable repository, dataset, or harness—not whether setup is easy.
Artificial Analysis
recommendedpartial
2026-10-26
Tests
Runs a fixed set of evaluations on commercial and open models and combines them into one Intelligence Index. It also measures output speed, latency, and price per token.
Scores public benchmarks, not models. It rates each benchmark on how well it separates models, how much headroom is left before saturation, and how much the field uses it.
Best for
Choosing which public benchmarks still carry useful signal.
Checks the quality of a benchmark with three metrics. These are ranking agreement with other benchmarks, power to separate models, and cases where a weaker model beats a stronger one in the same family.
Best for
Auditing whether an evaluation measures the capability its label implies.
HELM (Holistic Evaluation of Language Models) runs models on many scenarios with several metrics each. It publishes the prompts, metrics, and raw model outputs for each run.
Ranked open-weight large language models (LLMs) on a fixed set of public benchmarks. Hugging Face retired it in 2025, and its results now stand as an archive.
IFEval (Instruction-Following Eval) gives the model prompts with instructions a program can check, such as word counts or required keywords. It tests whether the output follows every instruction.
Asks the model to produce JSON or YAML output that follows a given schema, phrased in many different ways. A validator checks only the structure, not the content.
Best for
Comparing schema compliance before testing private structured-output contracts.
A hosted leaderboard that tests models on private agent and tool-use tasks. It reports a composite pass^5 score, which counts a task only if the model passes on all five tries.