LLM rankings in 2026: how leaderboards work and which to trust
By Felix Wunderlich -
Updated August 17, 2026. Vote counts, index versions and leaderboard mechanics checked August 2026.
TL;DR: LLM rankings come from three very different machines. Artificial Analysis runs its own standardized evaluations and publishes a weighted Intelligence Index, Arena (formerly LMArena) turns millions of blind human votes into Elo-style scores, and static benchmark suites score fixed question sets that saturate and leak over time. Each rewards different things and each can mislead on its own, so trust the intersection, and when you are picking a model to actually run, add the numbers scores leave out: price, speed and hosting. That operational layer is what our LLM leaderboard puts on top of the Artificial Analysis data.
Every few weeks a new model tops some leaderboard, and the leaderboards rarely agree with each other. That is not because any of them is broken, it is because LLM rankings are built by fundamentally different methodologies, and each one measures something different. This guide explains how the major ranking systems work, what each rewards and misses, and how to read them when you need to choose a model rather than win an argument.
Which LLM ranking systems matter in 2026?
| Ranking system | What it measures | How it works | Blind spots |
|---|---|---|---|
| Artificial Analysis Intelligence Index | Composite of nine independently run evaluations | AA runs every eval itself with standardized settings and pass@1 scoring | Text-only and English-only, no human preference signal |
| Arena, formerly LMArena | Human preference | Blind pairwise battles, Elo-style scores with confidence intervals | Documented gaming vectors, style can beat substance |
| Static suites (HLE, GPQA Diamond, MMLU-Pro) | Accuracy on fixed question sets | Published datasets, exact-match or multiple-choice scoring | Saturation, contamination, sensitivity to scaffolding |
| Provider launch numbers | Whatever the vendor chose to report | Vendor-run evals with vendor-chosen scaffolds | Not comparable across vendors |
| Opper LLM leaderboard | AA indices plus price, speed and hosting | Built on AA's indices, with per-route prices, measured speed and hosting overlaid | Frontier subset, inherits AA's scope |
How does the Artificial Analysis Intelligence Index work?
Artificial Analysis is the closest thing the field has to an independent measurement lab. Its Intelligence Index, at version 4.1.1, combines nine evaluations in four weighted categories: agentic work counts 34%, coding 24%, scientific reasoning 24% and general capability 18%. The individual tests include Terminal-Bench and SciCode for coding, Humanity's Last Exam and GPQA Diamond for scientific reasoning, and agentic suites like τ³-Banking.
What makes it credible is the process rather than the test list. Artificial Analysis runs every evaluation itself against internal dataset copies rather than taking vendors' self-reported numbers, with standardized settings across all models, single-attempt pass@1 scoring, and a 95% confidence interval on the index it estimates at under one point. That makes scores genuinely comparable across vendors, which almost nothing else in this space is. The honest limits, which they state themselves: the suite is text-only and English-language, and a measured index tells you nothing about whether humans actually like a model's answers.
How does Arena rank models, and can it be gamed?
Arena, the platform formerly known as LMArena and before that Chatbot Arena, rebranded in January 2026, answers the question measured evals cannot: which model do people prefer? In its battle mode two anonymous models answer the same prompt and you vote before the identities are revealed, and those votes become Elo-style ratings with confidence intervals. The scale is real, more than 7.7 million votes across 389 models on the text leaderboard alone, with separate boards for agents, web development, vision, image generation and search.
It is also the ranking system with the best documented gaming story. "The Leaderboard Illusion", an April 2025 paper from Cohere Labs and university collaborators, showed that Meta tested 27 private Llama 4 variants on the arena before release while disclosing results selectively, and that proprietary models were sampled into battles considerably more often than open-weight ones. Arena published a response disputing the framing and noting that pre-release testing is open to any lab, and both documents are worth reading before you treat an arena rank as ground truth. Votes also reward confident formatting as much as correctness, which is why the boards carry a style-control toggle; human preference is a signal worth having, just not the only one.
What do static benchmarks like Humanity's Last Exam measure?
Underneath both systems above sit static benchmarks, fixed question sets with published scoring. The current flagship is Humanity's Last Exam, 2,500 questions across more than a hundred subjects assembled by the Center for AI Safety and Scale AI from roughly a thousand expert contributors, built explicitly because frontier models had saturated the previous generation of tests. Alongside it sit GPQA Diamond, MMLU-Pro and dozens of narrower suites.
Static benchmarks are cheap and reproducible, and they age badly in two ways. Questions leak into training data over time, and scores swing with scaffolding, prompting and sampling choices, so the same model can post visibly different numbers in two labs. The clearest sign that the era of standalone static leaderboards ended: Hugging Face retired its Open LLM Leaderboard in March 2025 and archived it after evaluating more than 13,000 models over roughly two years, saying it "could encourage people to hill climb irrelevant directions" as capabilities shifted toward reasoning and assistants. Static tests still matter, but today they matter mostly as ingredients inside independently run composite indices.
Should you trust provider self-reported scores?
Launch-post numbers are the fourth leg of the ranking ecosystem and the shakiest one. They are produced by the vendor, on scaffolds and settings the vendor chose, usually on the benchmarks where the model shines, and none of that makes them lies, it makes them incomparable. The same model scored independently under standardized conditions routinely lands lower than its launch numbers, which is exactly why independent re-runs exist.
Provenance questions extend to benchmarks themselves. FrontierMath, a well-regarded hard math benchmark built by Epoch AI, turned out to be funded by OpenAI, which also had visibility into a large share of the problems and solutions, a relationship disclosed only in December 2024 when o3's record score on it was announced. Epoch said plainly that it made a mistake by not being more transparent, and maintains a holdout set. The takeaway is not that any particular score is wrong, it is that three questions are worth asking of every benchmark: who funds it, who runs it, and is there a holdout set.
How do you read a leaderboard for your use case?
A rank is a shortlist generator, not a decision, so the useful move is to pick the view that matches your workload. For coding, start from a measured coding index, Artificial Analysis builds its own from Terminal-Bench and SciCode, and add Arena's web-development board if you want a human preference signal on top. For agentic workloads, the agentic evaluations now carry the largest weight in the AA index for a reason, and tool-calling quality separates models more than raw capability does. For high-volume production use, intelligence per dollar matters more than intelligence, and a model a few points down a ranking is often several times cheaper and faster, which is what head-to-head comparisons are for. We also measured gateway latency rather than guessing, because routing overhead is part of the same calculation.
Then there is consistency, which almost no leaderboard captures because standard scoring is single-attempt. When we ran one deliberately simple logic question ten times against every model in our car-wash test, only 5 of 53 models passed all ten runs. A model that is right nine times out of ten can look identical to a fully reliable one in a single-attempt table, so whatever the rankings say, verify the shortlist on your own tasks before standardizing.
How does the Opper LLM leaderboard work?
Opper is our product, so read this section as the vendor describing its own tool and check the page itself. The LLM leaderboard takes the Artificial Analysis intelligence, coding, agentic and math indices, with attribution, and overlays the operational data score-only boards leave out: the price on each hosting route, measured output speed in tokens per second, and where each model can run, including routes with European hosting and zero data retention. It covers the frontier end of the 700+ models available on the gateway, and the full model directory leads with the same ranking by default. The honest limits mirror its source: it is a frontier subset, and it inherits Artificial Analysis's text-only, English-language scope.
As of August 2026 the top of the intelligence index belongs to Anthropic and OpenAI, with the newest Claude Opus and Claude Fable releases at maximum reasoning effort trading the top slots with GPT-5-series models at their highest efforts, and Qwen's Max line the strongest entry from outside the US. One reading tip: reasoning-effort variants rank as separate rows, so the same model can occupy several slots at different scores and prices. Scores rev with every index version, which is why this paragraph names model lines and prints no numbers; the live leaderboard is the source.
The top five on our leaderboard right now
A snapshot from the live leaderboard, mid-August 2026, ranked by Intelligence Index; the live page always carries the current standings, prices and latency:
- Claude Opus 5 (Adaptive Reasoning, Max Effort), Anthropic
- Claude Opus 5 (Adaptive Reasoning, Xhigh Effort), Anthropic
- Claude Fable 5 (Adaptive Reasoning, Max Effort, Opus 4.8 Fallback), Anthropic
- Claude Opus 5 (Adaptive Reasoning, High Effort), Anthropic
- GPT-5.6 Sol (max), OpenAI
LLM rankings FAQ
How are LLMs ranked?+
Through three main methodology families: independent measured evaluations like the Artificial Analysis Intelligence Index, where one organization runs standardized tests itself, human preference voting like Arena's blind pairwise battles, and static benchmark suites with fixed question sets. Provider launch numbers are a fourth source, run and selected by the vendor. Opper's LLM leaderboard ranks frontier models on Artificial Analysis data with price, speed and hosting alongside.
What is the best LLM right now, in August 2026?+
As of August 2026 the top of the Artificial Analysis Intelligence Index belongs to the newest Claude Opus and Claude Fable releases at maximum reasoning effort, with GPT-5-series models at their highest reasoning efforts close behind and Qwen's Max line the strongest entry from outside the US. The best model for you depends on the task and the budget, so check the live LLM leaderboard, which shows the current intelligence, coding and agentic rankings next to price and speed.
What is a frontier model leaderboard?+
A leaderboard restricted to frontier models, the current top tier of capability, instead of every model ever released. Frontier model rankings are more useful for choosing what to run today because they exclude models nobody would pick anymore. Opper's LLM leaderboard is a frontier model leaderboard: frontier LLMs ranked by intelligence, coding and agentic scores, with price, speed and hosting shown alongside.
Can LLM leaderboards be gamed?+
Yes, and it is documented. "The Leaderboard Illusion" paper showed Meta tested 27 private Llama 4 variants on the arena before release with selective disclosure, while Arena disputes that framing and notes pre-release testing is open to any lab. Voting boards also reward style, which Arena's style-control setting tries to correct for. Independent measured evaluations are harder to game directly but face training-data contamination instead, so the practical defense is to cross-check rankings built on different methodologies.
What happened to the Hugging Face Open LLM Leaderboard?+
Hugging Face retired it in March 2025 and archived the results after evaluating more than 13,000 models over roughly two years, saying it risked encouraging hill-climbing in irrelevant directions as capabilities shifted toward reasoning and assistants. Its role has largely passed to independently run composite indices like Artificial Analysis and to human preference boards like Arena.
Which leaderboard should I use for coding?+
Start with a measured coding index; Artificial Analysis builds its coding score from Terminal-Bench and SciCode, and the Opper leaderboard shows that coding ranking next to price and speed. Arena's web-development board adds a human preference signal for web output specifically. Then test the shortlist inside your own harness, because agentic coding depends on tool-calling quality as much as raw capability.
Do benchmark scores predict real-world performance?+
Only partially. Static scores suffer from contamination and saturation, and single-attempt scoring hides variance. When we ran one deliberately simple logic question ten times against every model in our car-wash test, only 5 of 53 models passed all ten runs. Treat rankings as a shortlist generator, then verify the shortlist on your own tasks.
How often do LLM rankings change?+
Every major release reshuffles the top, which in 2026 means monthly or faster, and the methodologies rev too, with the Artificial Analysis Intelligence Index at version 4.1.1. That is why hardcoded score tables age badly, and why this guide points at Opper's LLM leaderboard, built on Artificial Analysis data, instead of printing scores that would be stale within weeks.
The short version
Trust the intersection. A model that scores well on independently run evaluations, holds up in blind human preference, and behaves consistently on your own tasks is a safe pick, and any single leaderboard on its own is just one instrument. Start from the LLM leaderboard for the current frontier picture with price and speed attached, use compare when it comes down to two candidates, and browse the full model directory when your shortlist reaches beyond the frontier.