How scores are calculated.

GAIA leads because it is the broadest public agent benchmark with recognizable frontier difficulty. Terminal-Bench and tau2-bench stay close because they measure operational, tool-using task completion rather than pure QA. SWE Atlas, HiL-Bench, and the Remote Labor Index add narrower agent behaviors, but they remain supporting evidence instead of the backbone of the claim.

SWE-bench Verified stays on top because it is still the cleanest public proxy for real bug-fix work under repo constraints. LiveCodeBench and Aider stay near it because they measure contamination-resistant code generation and edit quality in active workflows. BigCodeBench, EvalPlus, and the agentic coding suites broaden coverage, but they are downweighted when they duplicate the same narrow pass/fail coding loop.

OSWorld-Verified leads because end-to-end computer-use task completion matters more than pointwise grounding alone. ScreenSpot-Pro remains important because visual grounding is a prerequisite skill, but it is still a sub-capability rather than the whole desktop task.

EQ-Bench v3, Creative Writing v3, and Judgemark all come from the same publisher and partially share evaluation machinery. I can describe what each sub-panel measures, but I cannot yet defend a public broad-EQ weighting as an independent ranking claim against a hostile reviewer. This vertical stays visible as source evidence until a second independent public EQ-style suite exists or the panel is narrowed to a claim I can defend.

withholding reason: weighting rationale not yet strong enough for a public winner claim

Three distinct failure modes, one per benchmark. Vectara HHEM targets document-grounded faithfulness — the failure mode most common in RAG pipelines where a model fabricates content that contradicts its source. SimpleQA tests closed-domain factual recall using short questions with unambiguous correct answers, exposed on OpenAI's own models but independently verifiable. TruthfulQA (new) isolates common misconceptions and conspiracy- adjacent claims that models absorb from pretraining data — a different failure mode from both faithfulness and factual recall. Together the panel covers the three most important hallucination axes; two of the three sources are fully independent.

BFCL Overall is weighted highest because it is the best broad proxy for production function-calling reliability across the suite. Multi-Turn and Live stay next because real systems break on stateful tool use and live tool execution, not just one-shot JSON. Non-Live remains useful, but it is the easiest part of the panel and overlaps heavily with the aggregate score.

Instruction following is a production-critical capability orthogonal to raw reasoning or writing quality — a model can write brilliantly while ignoring format constraints, word limits, or explicit negative rules. IFEval (Instruction Following Eval) is the cleanest public benchmark for this: 541 verifiable instructions (JSON output, word limits, forbidden keywords, language constraints) scored with no LLM judge — pure programmatic pass/fail. It is independently published by Google Research and run by both the Open LLM Leaderboard and LLM-Stats, giving two independent source publishers for confidence scoring. Panel deliberately stays narrow until a second benchmark with comparable rigor emerges.

Long-context capability is a distinct production requirement orthogonal to per-token quality: a model can score well on standard benchmarks while completely failing to synthesize information across hundreds of thousands of tokens. LongBench-v2 (THUDM/Tsinghua, 2024) is the strongest open benchmark for this: 503 bilingual, human-verified long-context questions across single-doc QA, multi-doc QA, long in-context learning, and code repo comprehension, with gold labels and no LLM-judge dependency — unlike NIAH and RULER which rely on synthetic retrieval tasks. Panel deliberately narrow while the field standardises: long-context evaluation methodology is still converging and most alternatives either lack verified data or cover too few models to be actionable.

AIME 2025 leads because it is the best public proxy here for hard, multi-step symbolic reasoning under real contest pressure. MATH-500 stays close because it covers broader math domains and solution styles. Open LLM Math and BenchLM Math broaden model coverage, while MMLU-Pro and Scale Math Evaluation stay lower as supporting evidence because they partially overlap with the harder dedicated math panels.

The broad multilingual claim is intentionally narrow: one aggregate row from each independent public suite, not every per-language slice multiplied into fake certainty. Global-MMLU-Lite gets the highest weight because it is the cleanest broad cross-language knowledge readout. LanguageBench adds diversity by mixing ARC, MGSM, and MMLU across languages. MGSM-avg is a dedicated multilingual math benchmark (10 languages) from OpenAI simple-evals — a useful third signal that catches language-math reasoning gaps invisible to knowledge-only suites.

Multimodal capability is a first-class differentiator between frontier models — most major models in 2026 are vision-capable and users routinely submit images, charts, and documents. MMMU and MMMU-Pro share top weight: MMMU provides the broadest coverage (57 academic subjects, strong model count) while MMMU-Pro uses vision-only inputs that prevent text-shortcut cheating, making it the harder and cleaner signal. MathVista isolates math reasoning with visual context (diagrams, charts, geometry) — a distinct capability from text-only math. MMStar is manually filtered to require genuine visual perception rather than language priors, guarding against contamination that plagues simpler VQA benchmarks.

The overall composite weights general instruction following and conversation quality highest, since these best predict whether a model is broadly useful. Reasoning and coding add depth. Agentic and tool use round out the picture for users who care about autonomous task performance.

GPQA Diamond is weighted highest because it is still the strongest public reasoning benchmark for novel, contamination-resistant expert questions. HLE and MMLU-Pro stay close because they stress frontier breadth and multi-domain rigor at larger scale. The rest are supporting panels that cover reasoning sub-modes such as legal analysis, factual recall, multimodal reasoning, and adversarial puzzle solving.

MCP Atlas gets the lead because it is the more direct benchmark for structured, multi-step tool orchestration in agent stacks. Toolathlon stays substantial because it adds breadth across tool scenarios, but it does not yet have the same practical signal weight as the stronger MCP-style workflow benchmark.

The uncensored panel is currently all one suite and mixes several normative ideas: willingness, broad intelligence, and free-form writing latitude. I can preserve the source evidence, but I cannot defend one public composite winner from this panel without importing value judgments that are not benchmark-stable. This vertical stays withheld.

withholding reason: weighting rationale not yet strong enough for a public winner claim

LMSYS Elo leads because large-scale pairwise preference still best captures whether humans actually prefer the prose. LiveBench Language and AlpacaEval stay close because they add fresher instruction-following and style quality checks. BenchLM, TutorBench, and UGI Writing remain supporting evidence that broaden writing coverage without overriding stronger direct preference and language-quality measures.