LLM Benchmark Comparison (Leaderboard Overview)

As of July 2026, no single model wins every LLM benchmark, but Anthropic's Claude Fable 5 tops the most: it currently leads MMLU-Pro (91.5%), SWE-bench Verified (95.0%), Humanity's Last Exam in both the no-tools and with-tools regimes, Chatbot Arena human-preference voting (1509 Elo), and the Artificial Analysis Intelligence Index (60).^{[1][3][4][6][11][14]} The remaining crowns split by capability: Google's Gemini 3.1 Pro leads the science exam GPQA Diamond at about 94%, and OpenAI's GPT-5.5 leads agentic Terminal-Bench and, with heavy test-time compute, the abstract-reasoning ARC-AGI-2 test (around 85%, alongside Google's ARC-Prize-verified Gemini 3 Deep Think at 84.6%).^{[2][7][9][13][17]} Several long-standing tests, MMLU-Pro, GPQA Diamond, and AIME 2025, are now effectively saturated at the frontier, so a one-point lead there is measurement noise rather than a real capability gap.^[2][5]

This page is a per-benchmark map: for each major evaluation it explains what the test measures, whether it is saturated, and which model currently leads, with the top few scores and a link to that benchmark's full page. It is a benchmark explainer, not an overall model ranking. For a single best-to-worst ordering of models and a survey of the ranking methods themselves, see LLM Rankings; for how these scores climbed over time, see the LLM Benchmarks Timeline; and for how the human-vote scores are computed, see the Elo rating system page.

Which model tops the most benchmarks in July 2026?

The short answer is Claude Fable 5. On independent leaderboards it is the number-one model on five of the benchmarks tracked here (MMLU-Pro, SWE-bench Verified, HLE, Chatbot Arena, and the Artificial Analysis Intelligence Index), and it is within a point or two of the lead on GPQA Diamond.^{[1][4][6][11][14]} That breadth, not a single headline number, is what sets it apart.

The other leaders divide along capability lines. Google's Gemini 3.1 Pro owns the science-knowledge crown on GPQA Diamond (94.1% to 94.3%).^[2][7] OpenAI's GPT-5.5 leads the two hardest agentic-reasoning tests, Terminal-Bench (paired with an agent scaffold) and ARC-AGI-2 (with unconstrained, expensive test-time compute).^[9][13] Open-weight models are close behind rather than absent: Zhipu's GLM-5.2, Moonshot's Kimi K2.6, and DeepSeek V4-Pro all cluster near 90% on GPQA Diamond and lead the open field, GLM-5.2 tops both the open-weight Artificial Analysis Index (51) and AIME 2026 (99.2%), and Google's Gemini models lead LiveCodeBench.^[1][2][10] The practical takeaway: choose the benchmark that matches your workload, because the "best" model changes with the task.

The LLM benchmark comparison table

Each row links to that benchmark's full page. "Saturated?" flags whether the frontier has compressed near the ceiling so the test no longer separates the top models. Scores are the best verified public results as of July 2026; where a vendor self-reports a number, an independent re-run (Artificial Analysis, Vals AI, Epoch AI) is preferred and labeled. Last verified: July 2026.

Scroll sideways for more →

Model links in the table point to each model's page where one exists; Claude Fable 5, Claude Sonnet 5, and GLM-5.2 do not yet have dedicated pages, so they appear as plain text. Scores marked "no tools" exclude web search and code execution; see the HLE and AIME sections for why that label matters.

What do MMLU-Pro and GPQA Diamond measure, and who leads?

MMLU-Pro is the harder successor to MMLU: 12,000 questions across 14 academic domains, each with ten answer options instead of four, which lowers the payoff of lucky guessing and rewards genuine reasoning.^[3][4] The frontier now sits around 89% to 91.5%, with the top handful of models inside about two points, so MMLU-Pro is approaching saturation while retaining a little separating power. On Vals AI's independent run, Claude Fable 5 leads at 91.5%, ahead of Gemini 3.1 Pro (91.0%) and Gemini 3 Pro (about 89.8% to 90.1%).^[4] Notably, OpenAI does not report an MMLU-Pro figure for GPT-5.5.

GPQA Diamond is the 198-question expert subset of GPQA, written by PhDs in biology, chemistry, and physics and filtered so that skilled non-experts with web access still score near chance.^[2] It is now heavily saturated: roughly ten models fall inside a four-point band from 90% to 94%, and because the set has only 198 questions, a one-point gap is about two questions, well within run-to-run noise. Gemini 3.1 Pro leads at 94.1% (Artificial Analysis) to 94.3% (Google's model card), with GPT-5.5 (93.5%), Claude Opus 4.8 (93.2%), and Claude Fable 5 (93.2%) just behind.^[2][5][7] Open-weight models are right there too, with GLM-5.2, Kimi K2.6, and DeepSeek V4-Pro near 90%.^[2] Human PhD experts score about 70% on the same questions, so the models are well past the expert baseline, and Anthropic has said it plans to stop reporting GPQA because it no longer discriminates.^[2][15]

Which model is best at coding? SWE-bench, Terminal-Bench, and LiveCodeBench

SWE-bench Verified is the most-cited agentic coding test: a 500-issue, human-validated subset of SWE-bench in which a model reads a real GitHub repository, writes a patch, and must pass the project's hidden tests.^[6][8] It is the clearest current example of a benchmark approaching saturation from below. On the independent Vals AI leaderboard Claude Fable 5 resolves 95.0% of issues, with Claude Opus 4.8 at 88.6% and GPT-5.5 at 82.6%; the open-weight and Google frontier (DeepSeek V4-Pro, Gemini 3.1 Pro, Kimi K2.6) clusters around 80%.^[6] Two cautions: SWE-bench numbers depend heavily on the agent scaffold, so vendor self-reports run higher than neutral re-runs (OpenAI reports 88.7% for GPT-5.5 versus 82.6% measured independently), and the field is already migrating to the harder SWE-Bench Pro as Verified tops out.^[6][8]

Terminal-Bench pushes further into open-ended agency: it drops an agent into a sandboxed shell and grades only whether multi-step tasks (compiling code, configuring servers, processing data) actually succeed, each attempted five times.^[9] It is not saturated. On the current curated Terminal-Bench 2.1 board the top pairing is OpenAI's Codex CLI with GPT-5.5 at 83.4%, just ahead of Claude Code with Claude Fable 5 at 83.1%.^[9] Every entry pairs a model with an agent harness, scores cannot be compared across the 1.0, 2.0, and 2.1 task sets, and a custom scaffold can outscore a standard agent on the same model, so read Terminal-Bench as a model-plus-harness result rather than a pure model score.

LiveCodeBench attacks contamination head-on. It continuously pulls fresh problems from LeetCode, AtCoder, and Codeforces, tags each with a release date, and scores a model only on problems published after its training cutoff, so memorized solutions cannot inflate the result.^[10] On the current v6 window the ceiling is about 91% Pass@1: Google's Gemini 3 Pro Preview leads Artificial Analysis's independent run at 91.7%, with Gemini 3 Flash (90.8%) and DeepSeek V3.2 Speciale (89.6%) close behind, and open-weight models (Qwen, Kimi K2.6) self-reporting around 89% to 92%.^[10] The companion LiveCodeBench Pro, which Elo-rates the very hardest contest problems, is nowhere near saturated: the best model sits near 2887 Elo against a top-human mark around 3800.^[10]

Which model is best at math and abstract reasoning?

AIME, the American Invitational Mathematics Examination, became a standard math benchmark in 2024: 15 questions with integer answers from 000 to 999, which makes it cleanly auto-gradable.^[16] Frontier reasoning models now score 100% on AIME 2024 and AIME 2025 with pure reasoning and no tools (a five-way-plus tie at the top), so those sets are fully saturated. The field has rotated to AIME 2026, which is near-saturated but still live: on MathArena's no-tools evaluation the leaders are open-weight, with Zhipu's GLM-5.2 reported at 99.2% and Kimi K2.6 at 96.4%.^[16] All AIME figures here are pure reasoning; some vendors quote a higher number with a code interpreter enabled, which is a different test and should not be mixed in.

ARC-AGI-2, from the ARC Prize Foundation led by Francois Chollet, is the benchmark most often misread. Its grid puzzles are easy for humans (about 60% average, 100% for a human panel) and were designed to be nearly impossible to memorize; at launch in March 2025 no public model exceeded single digits.^[13] That has changed at the high end: with heavy, expensive test-time compute the frontier has reached about 85%, led by Google's Gemini 3 Deep Think at 84.6% (the only score independently verified by the ARC Prize Foundation) and GPT-5.5 at a reported 85.0%.^[13][17] But ARC-AGI-2 is scored on a cost-versus-accuracy efficiency frontier, and its 85% grand prize (85% on the private set within a roughly $0.42-per-task budget, solution open-sourced) remains unclaimed: the 2025 Kaggle competition topped out at 24% under that constraint, and cheaper or non-reasoning models still score from the teens to the low 50s.^[13] So ARC-AGI-2 is simultaneously "almost solved" at any cost and "far from solved" efficiently. The genuinely unsaturated abstract-reasoning target is now the newer ARC-AGI-3, where models score under 1% and humans solve every task.^[13]

What are Humanity's Last Exam and the Artificial Analysis Intelligence Index?

Humanity's Last Exam (HLE) is the hardest broad public benchmark: about 2,500 questions written and vetted by experts across more than 100 subjects, built so that even frontier models fail most of them.^[11][12] It is nowhere near saturated. The single most important rule when quoting HLE is to never mix the two regimes, because a no-tools (text-only) score and a with-tools (web plus code) score are different tests. On Artificial Analysis's independent no-tools run over 2,158 text-only questions, Claude Fable 5 leads at 53.3%, ahead of Claude Opus 4.8 (45.7%) and Gemini 3.1 Pro (44.7%), with Kimi K2.6 the open-weight leader at 36.4%.^[11] With tools enabled, the best reported figure rises to about 64.5% for Claude Fable 5, but with-tools HLE is not a single standardized leaderboard (numbers come from vendor cards and reviews with differing tool setups), so those comparisons are looser.^[11] For reference, subject-matter experts score roughly 90% on HLE, which is the gap the benchmark exists to measure.^[12]

The Artificial Analysis Intelligence Index is not one test but a composite: version 4.1 (June 2026) aggregates nine independent evaluations spanning agentic tasks, coding, scientific reasoning, and general knowledge (for example GDPval, Terminal-Bench, SciCode, HLE, and GPQA Diamond), all run by Artificial Analysis rather than self-reported by the labs.^[1] Because the maintainers rebuilt the scale to undo the saturation of earlier editions, the current index tops out near 60 with real headroom below. Claude Fable 5 leads at 60, followed by Claude Opus 4.8 (56), GPT-5.5 (55), Claude Opus 4.7 (54), and Claude Sonnet 5 (53); GLM-5.2 is the top open-weight entry at 51.^[1] It is one of the better single numbers to cite precisely because it blends agentic, coding, and reasoning evals rather than rewarding one narrow skill.

Who wins on human preference (Chatbot Arena)?

Chatbot Arena, run by LMArena (the site moved from lmarena.ai to arena.ai in 2026), collects blind pairwise votes: a user sees two anonymous answers, picks the better one, and the votes are fit to a Bradley-Terry model that yields an Elo-style rating.^[14] As of July 1, 2026 the text leaderboard held 7.15 million votes across 369 models with style control enabled. Claude Fable 5 leads at 1509, and Anthropic's Claude family sweeps the entire top five (the thinking and standard variants of Opus 4.6 and 4.7 sit between 1494 and 1504), with Meta's Muse Spark, Gemini 3.1 Pro, and GPT-5.5 just behind.^[14] Two caveats matter: the top of the board is compressed within about 15 Elo, so rank order shuffles with each update, and Chatbot Arena measures preference, not correctness. A telling sign of the difference is that the more capable Claude Opus 4.8 sits below the older Opus 4.6 and 4.7 on human votes, a reminder that likability and raw capability are not the same axis.^[14]

Which benchmarks are saturated, and why it matters

Saturation is the central problem in reading any leaderboard. Once frontier models clear roughly 90% on a fixed question set, the remaining headroom is smaller than the noise between runs, and the test stops separating the best models. MMLU, GSM8K, and HumanEval saturated between 2023 and 2024; GPQA Diamond and AIME 2025 joined them in 2025 and 2026, and MMLU-Pro is close behind.^[2][5] The response has been to build tests that resist saturation: LiveCodeBench and its sibling LiveBench rotate in fresh, post-cutoff problems to defeat contamination; HLE is simply hard enough that the frontier is still near 50%; ARC-AGI-2 has been "solved" only at prohibitive cost, leaving its efficiency prize open; and the Artificial Analysis Index periodically rebuilds its scale. When a benchmark is marked "saturated" above, treat small differences as ties and weight the harder, less-saturated tests (SWE-bench, Terminal-Bench, HLE, and ARC-AGI-2's efficiency track) more heavily.

How this differs from our LLM Rankings page

This hub answers "what does each benchmark measure, and who leads it." For the complementary questions, use the neighboring pages: LLM Rankings surveys the ranking systems (LMArena, Artificial Analysis, Scale SEAL, LiveBench) and gives an overall model ordering; the LLM Benchmarks Timeline tracks how scores rose over time; and the individual benchmark pages linked in the table above go deep on methodology, history, and full leaderboards. Together they cover the ranking landscape without duplicating each other.

References

1. Artificial Analysis, "Artificial Analysis Intelligence Index" (v4.1 leaderboard and methodology), accessed July 2026. https://artificialanalysis.ai/evaluations/artificial-analysis-intelligence-index ↩
2. Artificial Analysis, "GPQA Diamond leaderboard," accessed July 2026. https://artificialanalysis.ai/evaluations/gpqa-diamond ↩
3. Artificial Analysis, "MMLU-Pro leaderboard," accessed July 2026. https://artificialanalysis.ai/evaluations/mmlu-pro ↩
4. Vals AI, "MMLU-Pro benchmark," accessed July 2026. https://www.vals.ai/benchmarks/mmlu_pro ↩
5. Vals AI, "GPQA Diamond benchmark," accessed July 2026. https://www.vals.ai/benchmarks/gpqa ↩
6. Vals AI, "SWE-bench Verified benchmark" (independent leaderboard, updated July 1, 2026), accessed July 2026. https://www.vals.ai/benchmarks/swebench ↩
7. Google DeepMind, "Gemini 3.1 Pro model card," accessed July 2026. https://deepmind.google/models/model-cards/gemini-3-1-pro/ ↩
8. SWE-bench, "Leaderboards," accessed July 2026. https://www.swebench.com/ ↩
9. Terminal-Bench (Laude Institute), "Leaderboard (Terminal-Bench 2.1)," accessed July 2026. https://www.tbench.ai/leaderboard ↩
10. LiveCodeBench, "Leaderboard" (release v6), and Artificial Analysis, "LiveCodeBench," accessed July 2026. https://livecodebench.github.io/leaderboard.html ↩
11. Artificial Analysis, "Humanity's Last Exam leaderboard" (no-tools, text-only subset), accessed July 2026. https://artificialanalysis.ai/evaluations/humanitys-last-exam ↩
12. Scale AI / CAIS, "Humanity's Last Exam leaderboard," accessed July 2026. https://labs.scale.com/leaderboard/humanitys_last_exam ↩
13. ARC Prize Foundation, "ARC-AGI Leaderboard" and "ARC Prize 2025 Results and Analysis," accessed July 2026. https://arcprize.org/leaderboard ↩
14. LMArena, "Text leaderboard" (arena.ai, formerly lmarena.ai; 7.15M votes, style control on), as of July 1, 2026. https://arena.ai/leaderboard/text ↩
15. Epoch AI, "GPQA Diamond," accessed July 2026. https://epoch.ai/benchmarks/gpqa-diamond ↩
16. MathArena, "AIME 2026" and "AIME 2025" (no-tools, average of 4 runs), accessed July 2026. https://matharena.ai/ ↩
17. Google, "Gemini 3 Deep Think: 84.6% on ARC-AGI-2, verified by the ARC Prize Foundation," February 2026. https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-deep-think/ ↩