The 2026 LLM Benchmark Reference - 17 Benchmarks with Current Scores

Knowledge Benchmarks

Knowledge benchmarks test breadth of factual understanding across academic domains. MMLU (2020) was the defining benchmark of the 2020-2023 era and is now saturated - frontier models score 92-94%, making it useless for current comparisons. MMLU-Pro addresses saturation with 10-choice questions and mandatory CoT, requiring more deliberate reasoning. Humanity's Last Exam (Scale AI, 2025) is the current frontier challenge: 3,000 questions by domain experts, designed to remain hard for several years. Current SOTA is around 18-20%, leaving enormous headroom.

MMMU adds multimodal understanding - questions require image interpretation alongside text. College-exam style across 30 subjects, it captures capability that text-only benchmarks entirely miss. For 2026 model comparisons, the recommended stack is: MMLU-Pro for text knowledge, MMMU for multimodal knowledge, and HLE for frontier reasoning limits.

Most useful in 2026: MMLU-Pro. Cite MMLU only for historical comparison pre-2024.

Code Benchmarks

Code benchmarks evolved rapidly through 2021-2026. HumanEval (OpenAI, 2021) with 164 Python function-completion tasks was the standard for years. It is now saturated: top models exceed 98% pass@1. MBPP (Google, 2021) with 974 basic Python problems is similarly exhausted. Neither benchmark should appear in a serious 2026 model comparison.

LiveCodeBench (2024) solved the contamination problem by drawing from LeetCode and AtCoder problems released after each model's training cutoff. This makes contamination structurally impossible rather than just unlikely. For 2026 code comparisons, use LiveCodeBench and SWE-bench Verified. SWE-bench Verified tests real-world software engineering capability - can the model write a patch for a real GitHub issue that makes failing tests pass? The gap between SWE-bench Verified scores and HumanEval scores reveals how different code completion is from software engineering.

Most useful in 2026: LiveCodeBench (anti-contamination), SWE-bench Verified (real engineering).

Reasoning Benchmarks

Reasoning is the benchmark frontier in 2026. GPQA (Rein et al., 2023) tests graduate-level, Google-proof scientific reasoning - questions written by domain PhDs that are specifically designed to be impossible to answer by searching the web. GPQA-Diamond, the 198-question hardest subset, is the canonical reference. Frontier models now score 71-78%, with GPT-5 reaching 78.4% as of April 2026.

ARC-AGI-2 (Chollet, 2025) is the successor to the original Abstraction and Reasoning Corpus. ARC-AGI was cracked in December 2024 when OpenAI's o3 scored 87.5% (above the 85% human baseline). ARC-AGI-2 introduces significantly harder visual abstraction tasks that resist the techniques that broke the original. SOTA is around 65.8% as of April 2026, with substantial headroom. BIG-Bench Hard is approaching saturation at 94.3%.

Most useful in 2026: GPQA-Diamond, ARC-AGI-2, Humanity's Last Exam. All have real headroom.

Agentic Benchmarks

Agentic benchmarks measure multi-step task completion with tool use - a fundamentally different capability from question answering. WebArena (2023) provides 812 tasks across real web applications (Reddit, GitLab, e-commerce) running in sandboxed browser environments. The agent must navigate, fill forms, search, and complete tasks without human assistance. SOTA is around 47%, suggesting substantial room for improvement and a real performance gap between frontier models.

AgentBench (2023) covers 8 agent environments including OS tasks, game playing, database management, and web browsing - 1,091 tasks in total. OSWorld (2024) focuses on computer use agents - can the model control a desktop application to complete a task? SWE-bench Verified (listed in code above but fundamentally agentic) is the most widely cited.

Most useful in 2026: SWE-bench Verified (coding agents), WebArena (web agents), OSWorld (computer use).

See agent benchmarks deep dive ->

Multimodal Benchmarks

Multimodal benchmarks have grown in importance as frontier models gained vision capability. MMMU (Yue et al., 2024) is the primary knowledge benchmark for multimodal models - college-exam questions that require understanding images alongside text, across 30 subjects. MathVista (2024) tests mathematical reasoning about visual content: geometry diagrams, data charts, and scientific plots. ChartQA and DocVQA cover document understanding.

MMMU has become a standard inclusion in frontier model cards in 2026 because it captures a capability dimension that text-only benchmarks entirely miss. GPT-5 leads at 81.2% as of April 2026, but the gap between models is narrowing.

Most useful in 2026: MMMU for multimodal knowledge, MathVista for visual math reasoning.

Human Preference Benchmarks

LMSYS Chatbot Arena uses blind pairwise human voting to build a Bradley-Terry Elo ranking. Over 2 million votes as of April 2026, making it the most statistically robust human-preference signal available. GPT-5 leads at Elo 1401, with Claude 4.5 Opus at 1389 and Gemini 2.5 Pro at 1371. The gap between the top three is within statistical noise for many query categories.

Chatbot Arena measures what humans prefer in open-ended conversation - a different dimension from correctness, coding ability, or agentic capability. A model can rank highly in Arena while performing poorly on SWE-bench, and vice versa. Both are valid measures of different capabilities.

Most useful in 2026: Chatbot Arena for preference, MT-Bench for instruction-following quality.

Human vs automated evaluation deep dive ->

Frequently Asked Questions