AI Blueprints
Agentic AI Suite
AirgapAI Products
Iternal AI Blueprint Builder
Enterprise AI initiative decision framework.
Evaluate AI opportunities across value, feasibility, cost, governance, risk, adoption, and implementation readiness.
Regulated & Public Sector
Commercial & Operations
AirgapAI Chat
Local, air-gapped LLM chat for classified environments.
Runs entirely on your own hardware for CUI, HIPAA, and other regulated workloads — no data leaves your network.
Explore AirgapAIGuides
AirgapAI Comparisons
Best-Of Guides
Benchmarks & Indexes
AI Strategy Blueprint
Free Chapter 1 (32 pages, PDF) — the framework behind 4 #1 Amazon best sellers.
Get your free chapterTraining by Role
Training by Industry
Learn & Research
On edu.iternal.ai
Iternal AI Academy
AI fluency for every role in your company.
912+ courses · 10-minute lessons · role-based paths.
Company
Partnerships
Contact
Work with Iternal
See the platform in a live demo.
Product walkthroughs for AirgapAI, Blockify, the AI Blueprint Builder, and the AI Academy.
Book a demo
AI Blueprints
AI Blueprint Builder AI Blueprint Books AI Strategy Guide (Hub)Strategy & Advisory
AI Strategy Consulting Fractional Chief AI Officer AI Executive Education AI Team Training AI AcademyAI Consulting Services
AI Consulting AI Implementation Services Machine Learning Consulting AI Data & Analytics Consulting Digital Transformation Consulting AI Agent Development Services AI Governance Consulting Generative AI ConsultingAgentic AI Suite
Turnkey AI Ultramemory Agent Skills Library AgentAuth Ultracache Iternal Workforce Waypoint RFP Blockify PRISM Nebulous IdeaFORGEAirgapAI Products
AirgapAI Chat AirgapAI Transcribe AirgapAI Code All PricingRegulated & Public Sector
Federal Contractors (CUI/CMMC) Defense & Aerospace State & Local (SLED) Healthcare & HIPAA Legal & Law FirmsCommercial & Operations
Financial Services Manufacturing Energy & Utilities Edge Deployment All IndustriesBy Role
CISOs & Security IT Leaders Compliance Officers Enterprise ArchitectsBooks & Partners
AI Partner Blueprint AI Strategy Blueprint Book AI Strategy Framework Free Partner Chapter For PartnersTools
AI ROI Calculators AI Readiness Assessment Browse All Assessments LLM Pricing CalculatorGuides
Guides by Job AI Cost & Budget Guide LLM Selection & Benchmarks Best Local AI Tools (Enterprise) On-Prem Hardware Sizing How to Deploy LLMs On-Premise Local LLM Deployment Guide Private LLM Guide AI Agent Security Checklist Agentic AI AI Agent Orchestration What Is an MCP Server? Token Usage & Cost Projections Reducing AI Token Costs How to Write AI PromptsLibrary
Case Studies Webinars Book a Demo Events What is Iternal? What is Blockify?AirgapAI Comparisons
vs Microsoft Copilot vs ChatGPT Enterprise vs Google Gemini vs Amazon Q vs Azure OpenAIBlockify Comparisons
All Blockify Comparisons Best Vector Databases Best RAG Frameworks Best Data Ingestion Tools Best AI GovernanceBest-Of Guides
Best Copilot Alternatives Best ChatGPT Alternatives Best Local AI Tools Best AI Books Best AI Consulting Firms Best AI Strategy Frameworks Top AI Companies in Austin Best Secure AI Transcription Best Private AI Coding Assistants Best Private AI Appliances Best Secure AI TranslationBenchmarks & Indexes
LLM Benchmark Repository All ComparisonsGet Started
Overview & Pricing Plans & Pricing AI-Fluency Quiz (2 min) AI Training for Teams Enterprise AI Training Guide Student Free AccessTraining by Role
For Marketing Teams For Sales Teams For Executives For Finance Teams For HR Teams For Legal Teams For OperationsTraining by Industry
For Manufacturing For Government For HealthcareLearn & Research
Best AI Training Courses AI Training for Employees How to Write AI Prompts AI Skills Gap AI Training ROIOn edu.iternal.ai
Browse All Courses Verify CertificateCompany
About Iternal Leadership CareersPartnerships
Partners AI Partner BlueprintContact
Contact Iternal Book a DemoFollow
LinkedIn TwitterAn LLM leaderboard ranks language models by benchmark score. This one publishes raw results — MMLU-Pro, GPQA Diamond, SWE-bench Verified, LiveCodeBench, Aider Polyglot, BFCL and Chatbot Arena ELO — for every major frontier and open-weights model, with token pricing on each row, source attribution on every score, and a daily refresh.
The full table is open: sort any column, filter to open weights, and jump to what each benchmark measures. Only the CSV export asks for an email.
| Provider | Model | Context | Input $/M | Output $/M | |||||
|---|---|---|---|---|---|---|---|---|---|
| Loading benchmark snapshot... | |||||||||
The most important LLM benchmarks in 2026 are MMLU-Pro (broad reasoning), GPQA Diamond (graduate-level science), SWE-bench Verified and Aider Polyglot (real-world coding), BFCL (tool/function calling), and Chatbot Arena ELO (human preference). Compare raw scores — not opaque composite indexes — against your specific use case, then validate the shortlist with our LLM selection guide.
The leaderboard above carries one column per benchmark. Each section below says what that benchmark tests, how the score is produced, and which source supplies it, with a link that reloads the table ranked by that column. Benchmarks whose sources have published no model scores yet are marked on their card instead of linking to an empty ranking.
These columns separate models on breadth of knowledge, factual reliability and multi-step reasoning. Together they are the closest thing to a general capability read, which is exactly why they belong side by side as separate numbers rather than averaged into one.
A multiple-choice exam across fourteen academic and professional domains. It extends the original MMLU with ten answer options instead of four and drops the items frontier models had already saturated, so the spread between models is readable again.
Rank every model by MMLU-ProThe hardest 198 questions of the Graduate-Level Google-Proof Q&A set, written by PhD holders in physics, chemistry and biology and validated so that non-experts with web access still fail them.
Rank every model by GPQA DiamondTwenty-three BIG-Bench tasks that language models had not beaten against the average human rater when the suite was published: logic grids, symbol manipulation, date arithmetic and multi-hop deduction.
Rank every model by BIG-Bench HardExpert-written closed-ended questions across mathematics, the natural sciences and the humanities, assembled by the Center for AI Safety and Scale AI to stay unsaturated after models cleared MMLU-style tests.
Rank every model by Humanity's Last ExamShort factual questions with a single verifiable answer, written by OpenAI so that a model either holds the fact or does not. It doubles as a hallucination read, because a confident wrong answer scores exactly as badly as no answer at all.
Rank every model by SimpleQAThe MMLU question set translated into fourteen languages. It separates models that carry their English knowledge across languages from models that lose a large share of it in translation.
Rank every model by MMMLUThe second Abstraction and Reasoning Corpus: coloured grid puzzles where each task carries its own rule that has to be inferred from two or three examples. It is built to resist pattern matching, so scores stay low even at the frontier.
Rank every model by ARC-AGI-2Math columns are the least forgiving on the leaderboard because the answer is either right or it is not. They are also where contamination shows up fastest, which is why the contamination-resistant sets belong in the same view as the classic ones.
Problems from the American Invitational Mathematics Examination. Every answer is an integer between 0 and 999, so scoring needs no judge model and no partial credit.
Rank every model by AIMEResearch-level problems commissioned by Epoch AI from working mathematicians and deliberately kept unpublished, so a model cannot have seen them during training.
Rank every model by FrontierMathThe hardest tier of the MATH competition set. Only Level 5 problems count, scored on the final answer rather than on the working.
Rank every model by MATH Level 5Multistep Soft Reasoning: long natural-language scenarios such as murder mysteries, object placement and team allocation, where the model has to hold several inferences at once to answer.
Rank every model by MuSRGrade-school arithmetic word problems that take a short chain of steps to solve. Saturated at the frontier, and kept on the leaderboard because it still separates small open-weights models from one another.
Rank every model by GSM8KThe coding columns are the most decision-relevant on this leaderboard because they are verified by test suites rather than by a judge model. Sort on the one that matches your workload: repository-scale patches, competitive problems, or multi-language edits.
Five hundred human-validated GitHub issues drawn from real Python repositories. A model has to produce a patch, and the patch only counts when the repository’s own test suite passes.
Rank every model by SWE-bench VerifiedCompetitive-programming problems published after each model’s training cutoff, on a rolling window. The design makes the score hard to inflate by memorising public solutions.
Rank every model by LiveCodeBenchExercism exercises across C++, Go, Java, JavaScript, Python and Rust, scored by whether the model’s edited files pass the exercise unit tests. It measures editing an existing file, not writing one from scratch.
Rank every model by Aider PolyglotOne hundred and sixty-four hand-written Python functions scored against unit tests on the first attempt. Saturated at the frontier, and kept on the leaderboard because it still separates smaller and older open-weights models.
Rank every model by HumanEvalAgent workloads fail on tool calls and long horizons rather than on knowledge. These columns are the ones to sort by before putting a model behind a function-calling API or a shell.
The Berkeley Function-Calling Leaderboard. Emitted calls are compared by abstract syntax tree against the expected call, so a model cannot score by describing the call in prose or by guessing plausible arguments.
Rank every model by BFCL V4End-to-end tasks in a real shell. The model runs commands, reads the output and recovers from its own failures before the task is scored complete, which makes it a long-horizon test rather than a single-turn one.
Rank every model by Terminal-BenchA continuously refreshed SWE-bench variant built from recent repository issues, so the task pool keeps moving ahead of training cutoffs.
Rank every model by SWE-bench LiveMulti-step tool-use problems with supervision at each step, so the score reflects where a chain of calls broke rather than only whether the final answer was right.
Rank every model by ToolCompCapture-the-flag security tasks that require running commands in a shell and chaining several steps before a flag comes back. It reads agentic execution under adversarial conditions rather than security knowledge.
Rank every model by CybenchInstruction following is what breaks structured output in production: the model is capable, and it still returns six bullets when the schema asked for four. Two of these three columns are checked by a program; the third is judged by a model and tagged accordingly.
Verifiable instruction constraints such as word counts, casing, formats and forbidden words, checked by a program rather than by a judge model.
Rank every model by IFEvalHeld-out constraint types that do not appear in IFEval-style training data, which separates genuine instruction following from memorised formats.
Rank every model by IFBenchReal user tasks collected from in-the-wild traffic and scored by an LLM judge against per-task checklists. It carries the JUDGE tag on this leaderboard because the scorer is itself a model.
Rank every model by WildBenchPreference ratings answer a different question from accuracy benchmarks: which answer did a person prefer when the two were shown side by side. Useful, and never comparable to a percentage score.
Crowdsourced pairwise votes between two anonymous models, converted into an ELO-style rating that generally sits between about 1,000 and 1,600. It is a preference signal, not an accuracy measurement, so it is tagged PREF and styled apart from every objective column.
Rank every model by Chatbot Arena ELOTwo more columns sit beside the benchmark scores: input and output price per million tokens, taken from the OpenRouter catalog in the same daily snapshot as the scores. Sort by either one to see the cost of the models a benchmark column just ranked, and read the pair together — an output-heavy workload can cost several times more on a model with an attractive input price.
What the price columns do not cover. Throughput and time-to-first-token are properties of the serving stack, not of the model: the same weights answer at different speeds on different hardware, batch sizes and quantisations, so this leaderboard does not publish a speed column it cannot source. Model the full bill with the LLM pricing calculator, size the hardware behind a self-hosted deployment with the hardware sizing guide, and shortlist against requirements in the LLM selection guide.
Almost every LLM leaderboard on the public internet publishes a composite intelligence index — a single 0–100 number computed by weighted-averaging whatever benchmarks happened to be available for each model. These composites look clean and sortable, but they have three fundamental problems that make them actively misleading for real decisions.
Problem 1: composites hide sparsity. Imagine two models. Model A scored 80 on GPQA and nothing else. Model B scored 80 on GPQA, 79 on MMLU-Pro, 78 on SWE-bench, 82 on MATH, and 77 on BBH. A composite "intelligence index" might give both models the same 80 — because the average of 80 is 80, and the average of (80, 79, 78, 82, 77) is also ~79. You cannot tell them apart from the composite alone, even though Model B's number is backed by 5× the evidence and Model A's is a single data point that could easily be wrong.
Problem 2: composites mix incomparable scales. Chatbot Arena ELO ranges from ~1000 to ~1600 and measures subjective human preference. MMLU-Pro is a 0–100% accuracy score on multiple-choice questions. SWE-bench Verified is the percentage of GitHub issues a model resolved that passed the actual test suite. These three things measure completely different phenomena and cannot be meaningfully averaged — but composites do it anyway.
Problem 3: composites hide what was weighted. Every composite has a proprietary formula. Is Intelligence weighted 30% or 40%? Does Coding include SWE-bench Live or only SWE-bench Verified? Does the "Reasoning" index include AIME, or not? You have to read a methodology page to find out — if one even exists. Raw benchmarks need no explanation: the number means exactly what the benchmark page says it means.
Raw scores also do not transfer one-for-one to enterprise workloads: a model that tops a public eval can still miss on your contracts, policies and product data. For that counterpoint, see why the most powerful model is not the most accurate on company data.
Many of the benchmarks from the GPT-3 era are now fully saturated — top models routinely score 95%+ on MMLU, HumanEval, and HellaSwag, making them useless for distinguishing frontier capability. The benchmarks that do matter in 2026 share three properties:
Use the open-weight filter on the table above to see how locally deployable models place against closed frontier systems, and pair it with the LLM rankings 2026 for a shortlist organised by deployment scenario.
The Intelligence and Coding champions on this leaderboard are increasingly open-weight models — Llama 4, DeepSeek V3.2, Qwen 3.6, Kimi K2 Thinking, MiniMax M2.7 — that match or beat GPT-5 on specific evals at 1/20th the cost. The challenge for enterprises is running them privately: fully air-gapped, with no data leaving the device.
Iternal's AirgapAI platform runs any of these open-weights models locally with Intel Core Ultra chipsets, requires zero cloud dependency, and matches GPU-class performance on modern laptops. Use this benchmark repository to pick the right model, then deploy it without giving your data to a hyperscaler.
The repository refreshes every 24 hours via a Cloudflare Cron Trigger. On each run, a worker fetches from four live sources in parallel: OpenRouter for the model catalog, pricing, and context windows; SWE-bench for software engineering scores; the LMArena community mirror for crowdsourced Chatbot Arena ELO; and the Hugging Face Open LLM Leaderboard v2 for MMLU-Pro, GPQA, BBH, IFEval, MATH Lvl 5, and MuSR. Benchmarks the live sources don’t cover — FrontierMath, HLE, LiveCodeBench, Aider Polyglot, BFCL V4, HumanEval, Terminal-Bench, AIME — are backfilled from a curated seed compiled from model-provider announcements and public research papers. The health of every source is shown in the summary strip above the table, and any single source failure simply falls through to the next without breaking the snapshot. Every cell in the detail drawer shows exactly which source provided the score and when it was fetched.
Pick your model here. Run it on your own hardware with AirgapAI — no cloud, no data egress, no per-token bills. Match GPU-class performance on Intel Core Ultra laptops.