Benchmark performance
Compare measured intelligence, cost, speed, and latency. Check the model variant and the units shown in each column.
Explore Artificial Analysis in a new tabTHE MODEL OBSERVATORY
Models Ranking.
Understand the scores. Compare the trade-offs.
Independent benchmarks, model specifications, and a practical cost comparison. Build your shortlist with the evidence in view.
Evidence before hype.No sign-in needed
SOURCE: ARTIFICIAL ANALYSIS
Language model leaderboard
Use the table’s filters and column headings to explore the results.
Opening the source view…
View blank or blocked? Open the leaderboard directly in a new tab.
Current source. Clear context. The time above records when the view opened, not when a model was evaluated. New results appear when the source publishes them. Refresh pauses in other views, offline, and while you interact inside the table. Turn it off to keep filters between refreshes; your preference is saved on this browser.
A score is the starting point.
Check capabilities and provider pricing before you choose.
Opening model explorer…
CHOOSE YOUR EVIDENCE
Different questions. Different leaderboards.
Compare the task, model version, and test setup. These sources measure different things and should not be blended into one score.
Human preference
Explore which responses people prefer in comparisons. Preferences can vary by task, language, and category.
Explore Arena in a new tabReal-world usage
See which models people use on OpenRouter. Popularity measures adoption; it is a different signal from benchmark performance.
Explore OpenRouter in a new tabSoftware engineering
Explore results on repository issues. Compare the benchmark split and agent setup as well as the underlying model; results measure the full system.
Explore SWE-bench in a new tabTHE DETAILS THAT MATTER
Read a model card with confidence.
- Benchmark score
- A result on a defined test set under a particular setup. Compare the same benchmark version, model variant, and reasoning settings. A higher score does not guarantee better results for every task.
- Preference and usage
- Human preference reflects which responses people choose. Usage reflects adoption on a particular service. Neither is interchangeable with accuracy or a benchmark score.
- Context and output limits
- Context is the amount of information a model can work with in one request. Output is the maximum response size. Limits are measured in tokens and may depend on the provider.
- Input and output pricing
- Input tokens are sent to the model; output tokens are generated. The explorer shows listed USD rates per million tokens. Caching, reasoning, images, tools, subscriptions, and tiered pricing can change the final bill.
- Open weights
- The trained model weights are available under the publisher’s terms. This does not by itself mean unrestricted open source, commercial permission, or free hosting.
- Unknown is not no
- A missing capability or price means the source has not supplied that information. Zero is an explicit listed rate, and may still involve access restrictions or subscription costs.
- Speed and latency
- Speed describes how quickly output is produced. Latency describes how long you wait for output to begin. Prompt length, load, routing, and reasoning settings can affect both.
- Source timestamps
- Retrieved time tells you when DataVility fetched the catalog. A source’s model update date refers to its metadata. Neither confirms a new benchmark evaluation or a current price quote.