LiveBench explained: the contamination-free benchmark that anchors our League Score.
LiveBench explained in plain terms: a benchmark built so its questions cannot leak into training data, scored across seven task suites. Here is what a LiveBench score measures, how the overall number is calculated, and how we turn it into the largest single component of the League Score.
By AI Pro League · Published
LiveBench carries 40% of the League Score on this site — more than any other single component — because it is the broadest general-capability test we track. It is also the benchmark whose headline number moves most often: our own roundup of the biggest AI model ranking changes traced the League’s largest one-day fall to a single LiveBench re-measurement. This explainer uses our own LiveBench leaderboard as the worked example throughout.
What LiveBench measures
LiveBench describes itself, on livebench.ai, as “a challenging, contamination-free long-horizon AI benchmark”. Contamination is the problem it is built to solve: a fixed test published once can end up inside a model’s training data, after which the score measures memorisation as much as capability. LiveBench instead scores models across seven task suites — agentic work, design, coding, browser use, data analysis, research and reasoning — so the overall number reflects long, multi-step tasks rather than short quiz questions alone.
How a LiveBench score is calculated
On LiveBench itself, the overall score is a weighted mean of the suites a model has a score in: the agentic suite counts double, design and coding count one and a half times, and the remaining suites count once. Our League Score then normalises that raw overall across the field we track — today the recorded range runs from 38.7 to 85.7 — and weights the result at 40%. The worked example, from the top of our table:
| Model | Raw score | Normalised (0–100) |
|---|---|---|
| Claude Fable 5.1 | 85.7 | (85.7 − 38.7) / (85.7 − 38.7) × 100 = 100.0 |
| GPT-6 Astra | 85.6 | (85.6 − 38.7) / (85.7 − 38.7) × 100 = 99.8 |
| Claude Opus 5.5 | 85.1 | (85.1 − 38.7) / (85.7 − 38.7) × 100 = 98.7 |
Because normalisation is relative to the field, a model’s component can move when a rival is re-measured, even if its own raw score never changed — exactly what happened across the table on October 26. Full method: how the League Score is built.
What a good LiveBench score looks like
31 of the 59 models in our database carry a recorded LiveBench score. The field is wide: the median recorded score is 69.8, the floor is 38.7 and the ceiling is 85.7. The top of the table:
| # | Model | Maker | Score | Measured |
|---|---|---|---|---|
| 1 | Claude Fable 5.1 | Anthropic | 85.7 | Oct 26, 2026 |
| 2 | GPT-6 Astra | OpenAI | 85.6 | Oct 26, 2026 |
| 3 | Claude Opus 5.5 | Anthropic | 85.1 | Oct 26, 2026 |
| 4 | GPT-6.1 Sol | OpenAI | 84.1 | Oct 26, 2026 |
| 5 | Claude Opus 5 | Anthropic | 82.1 | Oct 14, 2026 |
The full table, with a source link on every row, is the LiveBench leaderboard. The shortlist it feeds is our best AI chatbots page, and the standings it drives are the AI model rankings.
Why LiveBench scores move — and when to distrust a gap
A contamination-free benchmark has to keep its questions fresh, which means the test itself changes over time: a score recorded in one month is not measured against exactly the same question set as a score recorded later. That is why our table carries a measured date on every row, and why a re-measurement can drop a model sharply — on October 26 one model’s recorded score fell from 74.8 to 63.0 without the model itself changing at all. Treat small gaps between models measured on different dates as directional, not decisive.
The limits to keep in mind
- Coverage is partial. Just over half the models we track have a recorded score, so absence from the table means untested, not weak.
- The overall hides the profile. Two models with the same overall can be far apart on individual suites — a strong coder can be a weak browser user. Check the suite split before choosing for one kind of work.
- Scores age quickly. Fresh questions mean fresh measurements; a months-old score describes an older test as much as an older model.
- It is one component. At 40% of the League Score it is the largest single voice, but reasoning (GPQA Diamond), coding (SWE-bench Verified) and human preference (LMArena) still shape the final table.
For the companion tests, start with our explainers of GPQA Diamond and SWE-bench Verified, or the wider guide to AI benchmarks explained.