Every month brings new model releases, and every release claims to be the best. Static reviews go stale in weeks, so developers and curious users increasingly rely on AI model tracker tools: live leaderboards and indexes that rank models on real evidence. This guide compares the tools that actually matter, explains what each one measures, and shows how to combine them when you pick a model.
1. LMArena (formerly Chatbot Arena)
LMArena is the reference platform for ranking conversational AI models by human preference. Launched in 2023 by the LMSYS research collective at UC Berkeley, it shows you two anonymous model answers side by side, lets you vote for the better one, and converts millions of blind pairwise votes into an Elo-style ranking using the Bradley-Terry statistical model. In 2026 it has collected millions of votes across hundreds of models and now operates as an independent company at lmarena.ai, where it is simply called Arena.
What it measures: how good conversations feel to real people. It answers "which model feels better?" rather than grading models on exams.
Strengths: fully free to browse and vote; huge voter base smooths out noise; separate leaderboards for coding, long context, and other categories keep comparisons honest.
Weaknesses: it is a preference poll, not a benchmark — popularity, verbosity, and formatting can inflate scores. Single-digit Elo gaps are noise, so read it in tiers, not exact ranks. Only votes cast while both models are anonymous count toward the official standings, and the platform warns that conversations may be used for research, so do not paste secrets into battles.
Use it when: you care about everyday chat quality and want to see which models real users prefer.
2. Artificial Analysis
Artificial Analysis takes the opposite approach: instead of votes, it tracks measured performance. The platform monitors more than a hundred models with metrics including tokens per second, time to first token, cost per million tokens, and a quality score blended from multiple benchmark suites. It also covers image, video, and speech generation models, which most text-only leaderboards skip.
What it measures: the speed, price, and quality trade-offs of each model and provider.
Strengths: this is the tool you open when you need practical deployment answers. If your app needs responses in under two seconds, it shows which models and providers hit that target, and at what price.
Weaknesses: its quality scores are its own designed metrics and do not guarantee performance on your actual workload. Enterprise API access is quote-based, which makes cost discovery harder early in an evaluation.
Use it when: you are choosing a model API for production and need to balance latency, price, and quality.
3. OpenRouter Rankings
OpenRouter's rankings page is built from real API usage: which models developers actually route tokens through. It is a revealed-preference signal rather than an opinion poll — people pay per token, so usage reflects models that earn their keep in production.
What it measures: adoption. Token volume and request counts show which models the developer community is actually deploying.
Strengths: free to view; updated on a rolling window; hard to game in the way a vote-based board can be, because the signal costs money.
Weaknesses: usage reflects pricing and availability as much as quality — cheap, fast models dominate even when a frontier model is better. It tells you what people use, not what is best for your task.
Use it when: you want a reality check on which models are proven in production workloads.
4. LiveBench
Static benchmarks have a contamination problem: models may have seen the public test questions during training, so scores can be inflated. LiveBench fixes this by generating brand-new test questions every month. Categories cover reasoning, mathematics, coding, and instruction following, so scores reflect current capability rather than memorization.
What it measures: up-to-date reasoning, coding, and instruction-following performance on fresh questions.
Strengths: the monthly refresh makes it one of the most contamination-resistant public benchmarks available.
Weaknesses: it is still a fixed exam — a model that aces LiveBench can still write awkward prose or fail at your specific workflow.
Use it when: you want to track how models are improving month to month and verify that scores are not just memorized test answers.
5. Hugging Face Open LLM Leaderboard
The Hugging Face Open LLM Leaderboard was the community's standard reference for open-weight models, running fixed open-source test batteries and publishing results openly. It is worth knowing mainly as history now: it was archived in mid-2024 and retired around March 2025, so its rankings are frozen in time.
What it measured: static academic benchmark performance for open models.
Strengths: transparent, reproducible, and still a useful archive of where open models stood in 2024.
Weaknesses: dead data. Any model released after the archive date is absent, and older scores reflect an older era of model capability.
Use it when: researching the history of open models — never for current model selection. Treat any article citing it as its main source with suspicion.
6. SEAL Leaderboard (Scale AI)
The SEAL leaderboard from Scale AI uses expert human reviewers and private test sets that models have not trained on. The private datasets are the key design choice: they address the "teaching to the test" problem that plagues public benchmarks.
What it measures: expert-judged performance on unseen tasks.
Strengths: strong for enterprise evaluation, because the scores are less contaminated and the reviewers are domain experts rather than casual voters.
Weaknesses: smaller scale than crowd platforms, and the private datasets mean outsiders cannot fully audit the test set.
Use it when: you need a trustworthy shortlist for enterprise procurement and want scores that vendors cannot easily game.
7. Vellum AI Leaderboard
Vellum's leaderboard is aimed squarely at developers: it tracks platform-run evaluations and benchmarks designed for engineering decisions rather than marketing. It is more commercially oriented and covers a narrower model set than the big public boards.
What it measures: developer-relevant model performance, packaged for build-vs-buy decisions.
Strengths: practical framing; a free tier makes the basics accessible.
Weaknesses: narrower coverage, and the commercial orientation means you should cross-check its claims against an independent board.
Use it when: you are evaluating models inside an existing development workflow and want decision-ready numbers.
How to use them together: a practical workflow
No single leaderboard answers "which model should I use?" Each one answers a different question, so the smart move is triangulation:
1. Shortlist on preference: open LMArena and note the top tier in the category matching your use case — chat, coding, or long context. Do not sweat small Elo gaps.
2. Filter on cost and speed: take that shortlist to Artificial Analysis and check tokens per second, time to first token, and price per million tokens. Drop anything too slow or too expensive for your budget.
3. Sanity-check on reality: look at OpenRouter rankings to see whether developers are actually deploying your candidates, and at LiveBench to confirm their benchmark strength is current, not memorized.
4. Run your own test: put your finalists through 5–10 prompts from your actual work. In LMArena's side-by-side mode you can even test anonymously without affecting the public board. Your own task is the only benchmark that truly matters.
Pitfalls to avoid
- Chasing rank one. The difference between rank 1 and rank 4 is usually noise; the difference between the right and wrong price tier is real money.
- Quoting frozen boards. The retired Hugging Face leaderboard still appears in old blog posts. Always check the last-update date before trusting a ranking.
- Ignoring the contamination problem. If a leaderboard's tests are public, assume scores are partly optimized. That is why fresh-question benches like LiveBench exist.
- Forgetting the price of free. Voting platforms are free to browse, but anonymous battles may feed community research — keep proprietary prompts out of them.