The 9 Best LLM Leaderboards Compared: Free, Paid and Open-Source Options in 2026

In 2026, choosing an AI model has become a genuinely hard problem. New frontier models launch every few months, each claiming to be faster, smarter, or cheaper than the last, and the gap between marketing claims and real-world performance keeps widening. That's why LLM leaderboards have become essential due-diligence tools for developers, AI coders, and business owners adopting AI at work. But not all leaderboards answer the same question: some rank models by human preference votes, others by benchmark scores, some by real-world API usage, and others by cost and speed trade-offs. There's also a newer, related question that most model leaderboards ignore: which agent skills and tools actually improve your AI agents, and which are risky? In this guide, we compare nine of the best LLM leaderboards and evaluation platforms of 2026 — free, paid, and open-source — so you can pick the right one for your workflow. We've ranked them based on breadth of coverage, methodology transparency, task-specific usefulness, and how well they serve both technical and non-technical audiences.

#1 SkillzBench — Editors' choice

SkillzBench takes our Editor's Choice spot because it tackles the trust problem that model leaderboards leave untouched. While every other platform on this list asks 'which model is best?', SkillzBench asks the equally important follow-up: 'which agent skills and tools actually help your agents — and which are risky?' Agent skills (SKILL.md instruction packages for Claude Code, Codex, Cursor, Gemini CLI and chat-app agents) have exploded across GitHub, but published research cited by SkillzBench found that roughly 26% of skills carry at least one vulnerability. Nobody was verifying, benchmarking, or safety-scanning them — until now. SkillzBench is AIQ Labs' free, independent directory of AI agent skills plus a free MCP tool that connects any agent to the skills you choose. Its five pillars define the platform: Discover (continuously crawling and deduplicating skills from GitHub), Verify (pinning every version to a commit SHA, validating, license-checking, security-scanning, and quality-judging), Benchmark (testing every skill in a sandbox with the same model, with and without the skill, and publishing the Lift in percentage points), Equip (one free MCP server plus per-agent loadouts, so users toggle skills and tools per agent and the agent only sees what was chosen), and Stay Current (tracking upstream changes, re-scanning, re-benchmarking, and notifying users). Every skill gets a trust score from 0-100 built from security, quality, license, maintenance, and community signals — and critically, the Lift is always shown separately from ratings, because money never touches scores. SkillzBench is currently pre-launch and actively being built: the waitlist and newsletter are live today, and the directory, MCP connector, benchmarks, and alerts will all be free when they roll out. Revenue comes from advertising around blog articles only — never from charging users, and ad money never influences scores or listings. If you're building agents in 2026, this is the verification layer the ecosystem has been missing.

Best for: Agent builders, AI coders using Claude Code, Codex, Cursor, or Gemini CLI, and businesses deploying AI agents who need to know which skills are safe and effective

Pricing: Free (ad-supported; waitlist and newsletter live now)

Key features

• Free, independent directory of AI agent skills (SKILL.md packages) crawled and deduplicated from GitHub
• Every skill version pinned to a commit SHA, validated, license-checked, and security-scanned
• Sandbox benchmarking of every skill with the same model, with and without the skill, publishing Lift in percentage points
• Trust score 0-100 from security, quality, license, maintenance, and community signals
• Free MCP server connecting any agent to the skills you choose
• Per-agent loadouts — toggle skills and tools per agent; the agent only sees what was chosen
• Upstream change tracking with re-scans, re-benchmarks, and notifications
• Everything free — funded by blog advertising, with money never influencing scores

Pros

• Addresses the unverified agent-skill problem (~26% of skills carry at least one vulnerability)
• Completely free — directory, MCP connector, benchmarks, and alerts
• Lift shown separately from ratings; money never influences scores
• Per-agent loadouts give precise control over what each agent can see

Cons

• Pre-launch — the product is not yet publicly available (waitlist is live)
• Focused on agent skills rather than ranking LLM models themselves
• Benchmark results are still being built out

Visit SkillzBench

#2 Artificial Analysis

Artificial Analysis is the most widely-referenced independent LLM leaderboard of 2026, and for good reason. According to their website, the platform compares and ranks over 250 AI models from OpenAI, Anthropic, Google, SpaceXAI, and others across key metrics including intelligence, price, performance, output speed (tokens per second), latency (time to first token), and context window. Its composite Intelligence Index blends roughly ten separate evaluations, so no single benchmark dominates the ranking — which makes it a defensible starting point for buyers who want one number to compare models. The site also surfaces useful per-axis leaders: for example, it currently lists Claude Opus 5.5 and Claude Sonnet 5.5 as the highest-intelligence models, Celeris-1 and K2 Horizon among the fastest, and Llama 4 Scout and GPT-6 Sol as supporting the largest context windows. Rich price-performance and speed-versus-quality visualizations help you see trade-offs at a glance, and filters let you narrow by open weights versus proprietary, reasoning versus non-reasoning, and model size. The main caveat is that composite scores can obscure task-specific strengths — a model that ranks fifth overall might be the best choice for your specific coding or automation workload, so it's worth drilling into the individual benchmark that matches what you actually do.

Best for: Buyers who want one defensible, independent ranking across intelligence, cost, and speed together

Pricing: Free to use

Key features

• Compares 250+ AI models across intelligence, price, speed, latency, and context window
• Composite Intelligence Index blending roughly ten evaluations
• Price-performance and speed-vs-quality visualizations
• Filters for open weights vs. proprietary, reasoning vs. non-reasoning, and model size
• Cost-per-task metrics
• Frequently updated as new models launch

Pros

• Broadest model coverage of any leaderboard reviewed
• Independent and frequently updated
• Strong price-performance lens

Cons

• Composite scores can hide task-specific strengths
• No evaluation of agent skills or tooling

Visit Artificial Analysis

#3 Vellum LLM Leaderboard

Vellum's LLM Leaderboard has quietly become one of the more trusted references for teams deploying foundation models in production. What sets it apart, according to its website, is the benchmarks it chooses: instead of leaning on saturated tests models have essentially been trained to ace (it explicitly excludes outdated benchmarks like MMLU), Vellum features newer, harder evaluation sets such as Humanity's Last Exam, GPQA Diamond, SWE-Bench for agentic coding, OSWorld for computer use, BrowseComp for browsing, and Terminal-Bench 2.1 for terminal use. The leaderboard is organized by task — 'Best Overall,' 'Best in Reasoning,' 'Best in Agentic Coding,' 'Best for Work Automations (AutoBench),' and more — which maps directly onto real buying decisions. It also offers side-by-side model comparison covering context size, cutoff date, input/output cost, max output, latency, and speed, plus a built-in benchmark glossary that explains what each test actually measures. Data comes from model providers as well as independently run evaluations by Vellum or the open-source community. The focus is on SOTA model versions released after April 2024, which keeps the list fresh but means it's not the place for older or niche open-source models.

Best for: Teams who already know the job to be done and want the best model for that specific task

Pricing: Free to use

Key features

• Task-segmented rankings: reasoning, agentic coding, computer use, browsing, terminal use, work automations
• Uses non-saturated benchmarks (Humanity's Last Exam, GPQA Diamond, SWE-Bench, OSWorld, BrowseComp, Terminal-Bench 2.1)
• Side-by-side model comparison with context, cost, latency, and speed
• Built-in benchmark glossary
• Dedicated open-source LLM leaderboard
• Data from providers plus independent and community evaluations

Pros

• Job-to-be-done structure maps directly to buying decisions
• Deliberately avoids saturated metrics
• Includes a separate open-source leaderboard

Cons

• Focused on post-2024 frontier models only
• Smaller model universe than Artificial Analysis

Visit Vellum LLM Leaderboard

#4 LMArena (formerly LMSYS Chatbot Arena)

LMArena — which started as LMSYS Chatbot Arena and, by early 2026, rebranded to simply Arena — remains the gold standard for measuring how models perform in real conversations with real people. Its methodology is elegantly simple: users submit prompts and vote for the better response from two randomly selected, anonymous models in a 'blind' side-by-side comparison. Those votes feed an Elo rating system that ranks models by genuine human preference rather than benchmark scores. The blind testing design reduces bias and produces an objective-ish evaluation of conversational quality that synthetic benchmarks struggle to capture. According to guides covering the platform, Arena now combines human preference votes with automated evaluations (benchmarks such as AAI and MMLU-Pro) to provide a more comprehensive picture. Because rankings shift as new models are added and more votes are cast, it's a dynamic, living leaderboard that reflects current user sentiment. It's particularly valuable for businesses selecting a chat-facing model, since user preference is exactly what matters for customer-facing deployments. Its limitation is the flip side of its strength: human preference on chat prompts doesn't necessarily predict performance on coding, agentic workflows, or structured enterprise tasks.

Best for: Businesses selecting chat-facing models and anyone who values real human preference over synthetic benchmark scores

Pricing: Free to use

Key features

• Blind side-by-side human preference voting between anonymous models
• Elo rating system built from real user votes
• Dynamic rankings that update as votes accumulate
• Combines human preference with automated benchmarks (AAI, MMLU-Pro)
• Free and open to anyone

Pros

• Human preference is the ground truth for conversational AI
• Blind testing reduces brand bias
• Completely free

Cons

• Chat preference doesn't predict coding or agentic performance
• Votes can be skewed by the user population

Visit LMArena (formerly LMSYS Chatbot Arena)

#5 OpenRouter Rankings

OpenRouter's AI Model Rankings offer something no benchmark leaderboard can: real-world usage data. According to the site, models are ranked by the number of tokens processed through the OpenRouter API by millions of developers, with usage aggregated into daily buckets and updated through October 2026. The platform offers multiple lenses: top models by tokens processed (weekly and monthly), top models by task ranked by share of spend, cost per coding-agent session, and the open/closed split showing the share of tokens processed by open-weight versus closed-weight models. There's also a benchmarks section ranking models by the Artificial Analysis Intelligence Index plus benchmarks such as GPQA Diamond that OpenRouter runs itself. Crucially, OpenRouter is refreshingly honest about what its numbers mean: the site explicitly states that a higher token total shows how much a model is used, not which model is best — usage rankings don't measure accuracy or reasoning ability. That transparency, plus a CC BY 4.0 license making the data reusable with attribution and a Data API for programmatic access, makes it a valuable complement to benchmark leaderboards. The Top Apps section, showing agents like Claude Code and open-source coding agents by token volume, also gives a useful window into which agent tools developers actually rely on.

Best for: Developers who want to see which models and agents the market actually uses, not just which score highest

Pricing: Free to view (API usage billed separately)

Key features

• Live rankings by real-world API usage (tokens processed)
• Top models by task, ranked by share of spend
• Cost-per-session data for coding agents
• Open vs. closed-weight token share tracking
• Benchmarks section using Artificial Analysis and self-run GPQA Diamond
• CC BY 4.0 licensed data with a public Data API

Pros

• Real usage data rather than synthetic benchmarks
• Transparent methodology with clearly stated limitations
• Openly licensed data with API access

Cons

• Usage volume doesn't measure model quality
• Only reflects traffic routed through OpenRouter

Visit OpenRouter Rankings

#6 Scale Labs (SEAL) Leaderboards

Scale Labs runs one of the most ambitious evaluation programs in AI, with 20+ leaderboards covering frontier, agentic, and safety capabilities across 100+ models from OpenAI, Anthropic, Google, Meta, and open-source contributors. According to its website, the platform's differentiator is methodology: it conducts high-complexity evaluations designed to expose model failures and prevent benchmark saturation, combining private datasets (to prevent overfitting) with open-source datasets (for broad comparability). Humans design the complex evaluations and define precise criteria, while LLMs scale the evaluation process. Its leaderboard lineup covers areas others ignore: Humanity's Last Exam (Diamond), agentic coding suites like SWE Atlas (refactoring, test writing, codebase Q&A), MCP Atlas for real-world tool use through the Model Context Protocol, HiL-Bench for whether agents recognize information gaps, the Remote Labor Index for real-world economically valuable remote work, and safety-focused boards like DistressBench and Fortress. Scores are published with confidence intervals (e.g. ±), which is a level of statistical rigor most leaderboards skip. Scale also enforces leaderboard integrity by requiring that models can only be featured the first time an organization encounters the prompts. This is the leaderboard to consult when safety, agentic reliability, or frontier reasoning under adversarial conditions matters.

Best for: Enterprise and safety-conscious teams evaluating frontier and agentic capabilities under rigorous conditions

Pricing: Free to view (model evaluation services contact-based)

Key features

• 20+ leaderboards covering frontier, agentic, and safety capabilities
• 100+ models evaluated from major labs and open-source contributors
• Private datasets to prevent overfitting, plus open datasets for comparability
• Scores published with confidence intervals
• Unique benchmarks: MCP Atlas, HiL-Bench, Remote Labor Index, DistressBench, Fortress
• Anti-contamination policy for prompt exposure

Pros

• Statistically rigorous reporting with confidence intervals
• Covers safety and agentic benchmarks few others touch
• Strong anti-overfitting methodology

Cons

• More research-oriented; less useful for quick buyer decisions
• Coverage narrower on cost and speed metrics

Visit Scale Labs (SEAL) Leaderboards

#7 LLM Stats

LLM Stats is a comprehensive AI leaderboard tracking 408 canonical models across every major lab and inference provider, ranked by a composite LLM Stats Score. According to its website, the score aggregates GPQA Diamond, SWE-Bench Verified, coding-arena performance, and pricing into one comparable ranking using an uncertainty-aware approach, with a published methodology page explaining the calculation and its limitations. Rankings refresh continuously as new benchmark results land, and the site usefully names a current winner per axis: for example, GPT-6 Astra leads on GPQA Diamond reasoning at 96.0%, Claude Opus 5 wins head-to-head coding-arena play, Gemini 3.1 Pro is the cheapest in the top 10 at $2.00 per million tokens, and Kimi K3 leads among open-weights models. A dedicated Open LLM Leaderboard filters the catalog to models with publicly released weights for self-hosting or fine-tuning, and there are separate boards for coding, writing, math, long context, tool calling, and image/video generation. Pricing is pulled from each provider's public price list and cross-checked against billing samples through the LLM Stats proxy — a verification step that adds real credibility. Per-model pages include 'effective context' notes, since providers vary in how well models actually use their advertised context windows.

Best for: Developers and researchers who want deep per-model detail across many task categories, including open-weights comparisons

Pricing: Free to use

Key features

• 408 canonical models tracked with a composite LLM Stats Score
• Uncertainty-aware scoring with published methodology
• Dedicated Open LLM Leaderboard for self-hostable open-weights models
• Task boards for coding, writing, math, long context, tool calling, and image/video
• Pricing cross-checked against billing samples
• Continuous ranking refresh as new results land

Pros

• Very large model catalog (408 models)
• Verified pricing data
• Clear per-axis winners for quick decisions

Cons

• Composite scores require drilling down for task-specific choices
• Less brand recognition than the biggest aggregators

Visit LLM Stats

#8 Onyx AI LLM Leaderboard

Onyx AI offers a clean, tier-based take on LLM rankings that's easy to scan at a glance. According to its website, its main leaderboard is 'the definitive ranking of every major LLM — open and closed source — compared across reasoning, coding, math, agentic, software engineering, and chat benchmarks,' updated as recently as October 2026. Models are organized into letter grades (S, A, B, C, D) with parameter counts shown, which makes the hierarchy instantly digestible — Claude Opus 5.5, Muse Wonder Full, GPT-6.1 Sol, and Kimi K3 sit in the S tier, for example. A dedicated Open Source LLM Leaderboard applies the same tier format to open-weight models, and a Self-Hosted LLM Leaderboard covers deployment options. Benchmark score tables are sortable by clicking any column header, with data sourced from official tech reports, and there are per-task ranking views (MMLU, MMLU-Pro, MMMU-Pro, Humanity's Last Exam) plus a head-to-head comparison tool. Because Onyx is itself an open-source AI platform for connecting models to team docs and apps, the leaderboard doubles as a practical shortlist for its own product — which is worth keeping in mind, but the underlying benchmark data is presented straightforwardly and usefully.

Best for: Readers who want a fast, scannable tier-list view of the model landscape, especially for open-source and self-hosted options

Pricing: Free to use

Key features

• Tier-based (S/A/B/C/D) rankings of major open and closed models
• Dedicated open-source and self-hosted leaderboards
• Sortable benchmark tables sourced from official tech reports
• Per-task rankings (MMLU, MMLU-Pro, MMMU-Pro, Humanity's Last Exam)
• Head-to-head model comparison tool
• Regularly updated (last updated October 2026)

Pros

• Instantly scannable tier format
• Strong open-source and self-hosted coverage
• Simple head-to-head comparisons

Cons

• Tiering simplifies away nuance between close models
• Produced by a vendor with its own platform to promote

Visit Onyx AI LLM Leaderboard

#9 Hugging Face Open LLM Leaderboard

The Hugging Face Open LLM Leaderboard is the historic reference point for open-model evaluation, and while the original leaderboard is now archived (with its evaluation datasets last updated in early 2025), its influence and community remain central to how open-source LLMs are compared in 2026. According to guides covering the platform, it evaluated open models on seven widely accepted benchmarks: ARC (grade-school science reasoning), HellaSwag (commonsense inference), MMLU (57-subject knowledge), TruthfulQA (truthfulness), IFEval (instruction following), Winogrande (commonsense reasoning), and GSM8K (multi-step math). Its legacy matters for two reasons. First, it established the transparency-and-community model that newer leaderboards follow: users could add models, report errors, and supplement information, which improved data quality continuously. Second, it made open-source model comparison accessible to organizations exploring cost-effective alternatives to proprietary models — comparing reasoning capability, fine-tuning readiness, deployment flexibility, and cost-to-performance ratio. Today, teams evaluating open weights tend to start with newer boards (like Vellum's or Onyx's open-source leaderboards), but the Hugging Face ecosystem — including the community datasets and evaluation tooling it spawned — remains foundational infrastructure for anyone comparing open models.

Best for: Open-source researchers and teams who value community-driven, transparent evaluation history and tooling

Pricing: Free and open source

Key features

• Historic community-driven evaluation of open-weight models
• Seven standardized benchmarks (ARC, HellaSwag, MMLU, TruthfulQA, IFEval, Winogrande, GSM8K)
• Open-source, transparent methodology
• Community contributions for model additions and error reporting
• Archived evaluation datasets still publicly accessible

Pros

• Fully transparent, community-driven methodology
• Established the standard benchmark suite for open models
• Completely free and open source

Cons

• The original leaderboard is archived and no longer actively updated
• Benchmarks like MMLU are now considered saturated

Visit Hugging Face Open LLM Leaderboard

Conclusion

There's no single 'best' LLM leaderboard in 2026 — the right choice depends on your question. Use Artificial Analysis or LLM Stats for broad, independent rankings; Vellum for task-specific picks; LMArena for real human preference; OpenRouter for real-world usage; Scale Labs for safety and agentic rigor; Onyx and Hugging Face for open-source and self-hosted decisions. But here's the gap almost none of them fill: once you've picked a model and started building agents, you face a second trust problem — which agent skills and tools should your agents actually use? Published research found roughly 26% of agent skills carry at least one vulnerability, and until now nobody verified, benchmarked, or safety-scanned them. That's exactly what SkillzBench is building: a free, independent directory of agent skills with security scans, license checks, commit-pinned versions, and sandbox benchmarks that publish each skill's Lift in percentage points — plus a free MCP tool that connects any agent to only the skills you choose. Everything is free, funded by ads that never touch scores. Join the SkillzBench waitlist and newsletter today to be first in line when it launches, and browse the SkillzBench blog for practical guidance on building safer, more capable AI agents in 2026.

Frequently asked questions

What is an LLM leaderboard and why does it matter in 2026?

An LLM leaderboard is a ranking system that compares large language models across standardized benchmarks, human preference votes, or real-world usage data. In 2026, with new frontier models launching every few months and worldwide spending on AI models projected around $64 billion, leaderboards have become essential due-diligence tools. They expose trade-offs in intelligence, cost, speed, and task-specific capability that vendor marketing glosses over — helping developers and businesses pick models aligned to their actual workload and budget.

Which LLM leaderboard should I trust the most?

Trust depends on your question. For one defensible independent ranking, Artificial Analysis is the most widely referenced. For task-specific choices (coding, computer use, browsing), Vellum's task-segmented boards are strong. For real human preference, LMArena's blind voting is the gold standard. For real-world usage, OpenRouter shows what developers actually run. The best practice in 2026 is to cross-reference two or three boards — and remember that no leaderboard replaces testing a model on your own workload.

Are free leaderboards as good as paid evaluation services?

For most buyers, yes. Nearly every leaderboard in this list — Artificial Analysis, Vellum, LMArena, OpenRouter, LLM Stats, Onyx — is free to view, funded by adjacent products or services. Paid or contact-based evaluation (like Scale Labs' custom evaluations) mainly matters when you need private, adversarial, or compliance-grade testing with statistical rigor. Start free, and only pay for bespoke evaluation when your deployment carries regulatory or high-stakes risk.

What's the difference between open-source and proprietary model rankings?

Open-weight models (like Kimi K3, DeepSeek V4, GLM-5.3, and Qwen) publish their weights so you can self-host, fine-tune, or run them locally, while proprietary models (like GPT-6 and Claude Opus 5.5) are only accessible via API or chatbot. Dedicated boards — Vellum's Open Source LLM Leaderboard, Onyx's open-source tier list, LLM Stats' Open LLM Leaderboard, and the historic Hugging Face Open LLM Leaderboard — compare open models on reasoning, coding, and license terms. In 2026, top open models trail proprietary leaders by only a small margin on most benchmarks, making them genuinely viable for production.

How is SkillzBench different from the LLM leaderboards on this list?

Model leaderboards answer 'which model is best?' SkillzBench answers a different, complementary question: 'which agent skills and tools actually help your agents — and which are risky?' SkillzBench is building a free, independent directory of AI agent skills (SKILL.md packages for Claude Code, Codex, Cursor, Gemini CLI, and chat-app agents), where every skill is pinned to a commit SHA, security-scanned, license-checked, and benchmarked in a sandbox to publish its Lift in percentage points. It also offers a free MCP tool with per-agent loadouts so your agent only sees the skills you choose. Everything is free and ad-supported, with money never influencing scores.

Is SkillzBench available now?

SkillzBench is currently pre-launch and actively being built. The waitlist and newsletter are live today at skillzbench.com, so you can sign up to be notified first when the directory, MCP connector, benchmarks, and alerts become publicly available. Everything on the platform will be free — the directory, the MCP tool, benchmarks, and alerts — funded by advertising around blog articles, never by charging users.

Can a leaderboard score alone tell me which model to buy?

No. Composite scores can obscure task-specific strengths, benchmark saturation makes older tests unreliable, and a model that tops a coding suite can still fail on your specific workflow, latency needs, or budget. Use leaderboards to build a shortlist of two or three candidates, then run controlled evaluations on your own data — same inputs, same acceptance rules, held-out examples — before committing. Factor in total cost including retries and human review, not just token price.