Agent Benchmark Leaderboard 2026: AgentBench, SWE-bench, GAIA
17 benchmarks. 8 eval tools. What each measures, where it breaks, and which to quote. Every score we reprint is dated to its source.
Independent reference for ML engineers choosing, defending, and shipping models. Covers public leaderboards (MMLU-Pro, GPQA-Diamond, ARC-AGI-2), agent benchmarks (SWE-bench Verified, WebArena, Tau-Bench, OSWorld, Terminal-Bench), and the practitioner stack. Where a benchmark has a current, checkable primary leaderboard, the picker reprints its top rows with the date verified, the source, and the caveats that make them comparable; where it does not, we link the official board rather than freeze a number that moves weekly.
Which model wins for YOUR use case?
Pick what you're actually trying to build. We show the benchmark that matters and what its leaderboard says, re-verified against the cited source. Where no trustworthy per-model leaderboard exists, we say so instead of printing one.
Every leaderboard row above was re-verified against the cited primary source on 12 June 2026 (official SWE-bench data file, EvalPlus results.json, NVIDIA RULER repository). Tabs without rows name the reason. Disagreement with vendor marketing is expected and a feature.
Three Layers of AI Measurement
Most listicles cover the first layer in isolation. We thread all three: public benchmarks, agent benchmarks, and your own eval pipeline.
Public Model Benchmarks
The leaderboards everyone quotes. We cover 17 benchmarks across knowledge, coding, reasoning, and multimodal, with task count, release year, saturation status, and the official leaderboard link per row.
Agent Benchmarks
The newer, less-settled frontier. Multi-step task completion in real environments. Methodology is disputed and scores are still moving fast. We document the strengths and gaming vectors of each.
Your Own Evals
How teams actually measure the workflow they ship. Custom golden datasets, LLM-as-judge, online monitoring, regression tracking. Honest comparison of the practitioner stack, no vendor capture.
Coverage Map
We index 17 benchmarks across six categories. Each entry carries category, release year, task count, status (active, saturated, deprecated), and the official leaderboard link. Click any category to read the full reference.
Open the full reference →MMLU, MMLU-Pro, MMMU, HLE
HumanEval, MBPP, LiveCodeBench
GPQA, ARC-AGI, BIG-Bench Hard
SWE-bench, WebArena, Tau-Bench
MMMU, MathVista, ChartQA
Chatbot Arena, MT-Bench
The 2026 Frontier, In Tiers
Ranks change weekly; absolute scores change quarterly. We summarise tiers and trends rather than freezing a leaderboard that will be wrong by next month.
Per-Benchmark Deep Dives
Domain-Specific Benchmarks
Eval Methodology and Tools
Per-Framework and Per-Workload
Read Every Score Like a Reviewer
A six-question rubric we apply to every benchmark score we read. Borrow it for your own reading.
- 01
What is the capture date?
Frontier benchmarks move weekly, so a 2024 score is historical, not current. Ask it of every score you quote, and take the live number from the benchmark's official board rather than a reprint.
- 02
What N-shot, what CoT?
0-shot, 5-shot, and CoT-required runs are different tests. Quote the setting alongside the score or do not quote it.
- 03
Which test set version?
MMLU vs MMLU-Pro, SWE-bench vs Lite vs Verified, GPQA vs Diamond. The names are similar; the tests are not.
- 04
Vendor card or third party?
First-party model cards optimise for their own model. HuggingFace and Papers With Code are independent. Both have failure modes.
- 05
Public test set?
If yes, ask about contamination. MMLU questions appear in Common Crawl. HumanEval problems echo LeetCode. Verified subsets help, do not eliminate, the issue.
- 06
What harness, what tools?
Agentic scores depend on the harness. Best-of-16 with extended tools is not comparable to greedy single-shot.
The Practitioner Stack
Eight evaluation platforms reviewed independently. Open source, cloud, hybrid. Honest where the open-source option is good enough, honest where it is not. No vendor wrote this comparison.
Read the full review →What Most Listicles Miss
The benchmark landscape in 2026 has three systemic problems that most coverage ignores. First, contamination. MMLU test questions appear verbatim in Common Crawl. HumanEval problems are near-duplicates of LeetCode solutions in pre-training data. A 94% score on a saturated benchmark might reflect memorisation as much as reasoning, and there is no clean way to tell from the leaderboard.
Second, saturation. MMLU, HumanEval, and MBPP no longer discriminate frontier models. The field has moved to MMLU-Pro, GPQA-Diamond, ARC-AGI-2, and Humanity's Last Exam. Most comparison sites still quote the saturated versions because the numbers are larger and more familiar to their readers.
Third, methodology opacity. “Best-of-16 with chain-of-thought and tool use” is not comparable to “greedy zero-shot.” Scores published without methodology are unfalsifiable claims. Every table on this site documents the evaluation setup so the comparison stays honest.