benchmarkingagents.com is an independent reference for the public benchmarks used to evaluate large language models and AI agents in 2026. Coverage spans the 17 most-cited benchmarks across knowledge (MMLU, MMLU-Pro, MMMU, Humanity's Last Exam), coding (HumanEval, MBPP, LiveCodeBench, SWE-bench Verified), reasoning (GPQA, GPQA-Diamond, ARC-AGI, ARC-AGI-2, BIG-Bench Hard), agentic capability (SWE-bench Verified, WebArena, AgentBench, OSWorld, Terminal-Bench, Tau-Bench), multimodal (MMMU, MathVista, ChartQA), and human preference (LMSYS Chatbot Arena, MT-Bench).

It also covers the practitioner stack that production teams actually use to evaluate their own systems: custom golden datasets, LLM-as-judge methodology, RAG evaluation frameworks like Ragas, online production monitoring, and the eight evaluation platforms in widest use (Braintrust, Langfuse, LangSmith, Arize Phoenix, DeepEval, Patronus, Helicone, PromptLayer).

It is not a leaderboard that scrapes scores without context. Where a benchmark has a current, checkable primary board, the homepage picker reprints its top rows with the date verified, the source, and the caveats (scaffold, evaluation mode, saturation) that any quoted score depends on; where no trustworthy board exists, it says so rather than inventing a table, and the benchmark page links the official leaderboard for the live number. Where a benchmark has saturated or has documented contamination risk, the page says so explicitly.

The benchmark landscape has three structural problems that most AI coverage ignores or treats as footnotes. First, contamination. MMLU test questions appear verbatim in Common Crawl, the primary web-crawl dataset behind most pre-training corpora. HumanEval problems are near-duplicates of LeetCode solutions that pervade GitHub and Stack Overflow crawls. SWE-bench issues have solutions in the same repositories' public git history. A high score on a contaminated benchmark may reflect memorisation as much as reasoning, and the leaderboard cannot tell you which.

Second, saturation. MMLU saturated in 2024. HumanEval, HellaSwag and WinoGrande saturated in 2023 or 2024. All frontier models now score within statistical noise of each other on these benchmarks. The field moved to MMLU-Pro (10-choice, CoT-required), GPQA-Diamond, ARC-AGI-2 and LiveCodeBench specifically because the previous generation no longer discriminates. Most comparison sites still quote the saturated versions, because the numbers are larger and more recognisable.

Third, self-reporting. Model cards are published by the same company that built the model. Common issues: publishing only benchmarks where the model performs well, omitting evaluation setup (N-shot, CoT, harness, test set version), and reporting scores under maximally favourable conditions that independent replications struggle to reproduce. The benchmark-literate reader treats vendor-published scores as a starting point, not a final answer.

This site exists to apply a consistent six-question rubric to every benchmark score it cites, document the methodology behind every number, and refuse to publish vendor-favourable shortcuts when the underlying signal is noisy.

Independent reference. No vendor affiliation. No paid placements on the tool-comparison pages. No sponsored entries in any benchmark or evaluation framework cited. Provider order in tables is determined alphabetically or by category, not by any commercial relationship.

Outbound links to evaluation tool vendors on the /tools-compared page go directly to each vendor and are not monetised today. If an affiliate link is ever added it will be labelled, and it would not influence the order tools appear, the rubric used to compare them, or which tools are recommended for which use case. The neutral editorial framing holds regardless.

No display advertising. No newsletter capture. No content sponsorships. No vendor lead-routing. The site is a reference, not a lead-generation funnel.

SOTA benchmark scores are re-verified against the official leaderboard for each benchmark, prioritised by search position: the homepage task-to-benchmark picker and the agent-benchmark cluster heads (SWE-bench Verified, WebArena, AgentBench) first, the long tail after. The most recent sweep, covering the picker's six cited sources, closed in June 2026. Where a cited leaderboard turns out not to publish per-model rows, the page says so rather than reprinting numbers we cannot trace.

The verification date is held in a single constant (LAST_VERIFIED_DATE) in src/lib/schema.ts. Footer text, schema dateModified, and visible headings all read from that single source. This is a deliberate design choice so cosmetic-refresh leaks (rolling a date forward without doing the underlying verification work) are structurally prevented.