Live public datasetLatest source data: Sep 1, 2026Checked: September 3, 2026

AI model benchmark tracker.

A sortable view of frontier model performance across public Arena evals, with space for independent intelligence, cost, and coding benchmarks as the data layer expands.

View tableWhat These Mean
191

Models tracked

5

Arena views

22

Labs covered

0

SWE-bench matches

Current leader

Claude Fable 5

Text Arena

1507Rank #1

Code Arena

1628Rank #8

Vision Arena

1313Rank #1

SWE-bench

Not listed

191 models shown

Arena and SWE-bench scores are higher-is-better. Latency and price are lower-is-better.

Benchmark guide

What the scores mean.

A quick reading key for comparing models without over-interpreting small differences or missing source data.

Higher: Arena + SWE-benchLower: price + latency
What do the Arena columns measure?

Text, Code, Vision, Search, and Docs are separate public Arena views. They compare models in head-to-head preference evaluations for the specific task type, so scores should be read within each column rather than as one universal intelligence number.

How should I read score, rank, CI, and votes?

The main score is the leaderboard rating and the rank is that model's position in the same Arena view. CI is the confidence interval around the score, and votes show the amount of comparison data behind it. When confidence intervals overlap, small rank gaps deserve extra caution.

What does SWE-bench mean?

SWE-bench tracks how often a model solves real software-engineering issues from GitHub repositories. The table uses matched benchmark rows for model-level comparison, with higher percentages indicating more verified tasks solved.

What do price, latency, and speed mean?

Price is the estimated blended cost per one million tokens. Latency is time to first token, so lower is better. Speed is output throughput after generation begins, so higher is better.

Why are some benchmark cells empty?

An empty cell usually means the model is not listed in that source, the benchmark has not published a comparable score, or the public model name could not be matched confidently. It should not be treated as a zero score.