Language Model Benchmarks¶
How do you know whether one artificial intelligence (AI) model is better than another? Benchmarks are the field's answer: standardized test sets that every model takes, so results can be compared. This page covers language models: how to read their leaderboards, the ones worth knowing, and a live snapshot of one of them. Generative media is ranked differently; see the companion pages on image generation and video generation.
How to Read a Leaderboard¶
The Five-Minute Version
Five things to keep in mind when reading benchmark numbers:
- Benchmarks measure the test, not your task. A model that tops a math benchmark may still write mediocre patient-education materials. Scores are a compass, not a verdict; the only benchmark that truly matters is a trial on your own work.
- Contamination inflates scores. When a benchmark's questions leak into training data, models can score well by memory rather than ability. Newer benchmarks fight this by replacing their questions on a schedule or keeping some of them private. LiveBench, featured below, does both: it swaps in new questions periodically and keeps its newest questions unpublished.
- Small gaps are noise. Many leaderboards print an uncertainty range beside each score, such as a ± figure or a range of possible ranks. When two models' ranges overlap, treat them as tied. Where no range is shown, a point or two on a 100-point scale is rarely meaningful. Pay attention to tiers, trends over months, and category strengths (a model can be strong at coding and middling at instruction following), not single-rank differences.
- Weigh cost alongside score. Several leaderboards now plot score against cost and draw a frontier line through the models that no other model beats on both. A model below and to the right of that line costs more for a lower score. Check what "cost" means on each chart (price per million tokens, cost per task, per image, or per minute of video), and notice when the cost axis is logarithmic, where one step can mean several times the price. A model that writes long reasoning can cost more per task despite a lower price per token.
- Check the version and setting. The same model can appear several times at different reasoning-effort settings, each with its own score and cost, so check which setting a row describes.
For the vocabulary, see benchmark, eval, and leaderboard in the glossary.
The Leaderboards Worth Knowing¶
-
Artificial Analysis
Independent measurements of intelligence, speed, and price across providers, in a single view of cost against capability.
-
LiveBench
An open, contamination-aware benchmark that refreshes its questions on a schedule and reports what each model costs per successful task. The table below is refreshed automatically from its published data.
-
BenchLM
Gathers published results from hundreds of benchmarks, many reported by the model makers themselves, into one comparable view, and labels how much evidence stands behind each ranking.
-
Arena (Formerly LMArena)
Rankings from millions of blind head-to-head votes by real users, a measure of preference rather than test performance. Its text board can also plot score against price.
LiveBench Snapshot¶
LiveBench release 2026-09-02, fetched October 15, 2026. Scores are 0 to 100; the global average is the mean of the category averages. Data: LiveBench, an open, contamination-aware benchmark.
Where the top 15 differ
one model ranked first overall (claude-fable-5.1) lowest to highest
Overall (global average)74.9 to 88.0, a spread of 13.1 points
Long-Horizon40.4 to 85.0, a spread of 44.6 points
Reasoning86.7 to 95.1, a spread of 8.4 points
Coding77.3 to 98.3, a spread of 21.0 points
Browser72.0 to 100.0, a spread of 28.0 points
Data Analysis59.5 to 85.2, a spread of 25.7 points
Research54.5 to 88.9, a spread of 34.4 points
Design73.7 to 86.9, a spread of 13.2 points
| # | Model | Global | Long-Horizon | Reasoning | Coding | Browser | Data Analysis | Research | Design |
|---|---|---|---|---|---|---|---|---|---|
| 1 | claude-fable-5.1 | 88.0 | 80.8 | 91.7 | 97.7 | 100.0 | 80.3 | 82.0 | 83.4 |
| 2 | claude-opus-5.5 | 86.5 | 85.0 | 92.2 | 96.3 | 80.0 | 80.3 | 85.0 | 86.9 |
| 3 | gpt-6-astra | 86.1 | 82.4 | 92.7 | 93.7 | 84.0 | 83.0 | 85.2 | 82.0 |
| 4 | gpt-6.1-sol | 85.4 | 73.0 | 92.6 | 96.0 | 88.0 | 82.7 | 83.5 | 82.1 |
| 5 | gemini-4-argon | 84.1 | 69.9 | 95.1 | 89.7 | 92.0 | 85.2 | 83.1 | 73.7 |
| 6 | claude-fable-5 | 83.6 | 53.2 | 89.7 | 98.3 | 92.0 | 80.5 | 86.9 | 84.7 |
| 7 | claude-sonnet-5.5 | 82.5 | 78.8 | 91.6 | 94.7 | 84.0 | 59.5 | 83.6 | 85.0 |
| 8 | claude-opus-5 | 82.1 | 62.1 | 91.2 | 90.8 | 84.0 | 74.6 | 88.9 | 83.0 |
| 9 | gpt-6-sol | 81.9 | 66.9 | 88.7 | 93.1 | 92.0 | 81.2 | 69.3 | 82.2 |
| 10 | muse-wonder-1.0 | 78.1 | 53.2 | 90.6 | 86.2 | 72.0 | 81.6 | 88.0 | 75.1 |
| 11 | deepseek-v4.1-flash | 77.3 | 43.8 | 86.7 | 85.4 | 83.3 | 79.3 | 85.3 | 77.0 |
| 12 | claude-opus-4.8 | 76.8 | 48.3 | 89.2 | 77.3 | 96.0 | 66.0 | 81.8 | 78.8 |
| 13 | gpt-5.6-sol | 76.3 | 48.5 | 91.7 | 89.7 | 88.0 | 79.8 | 54.5 | 81.6 |
| 14 | gpt-5.6-terra | 75.0 | 48.4 | 90.6 | 86.2 | 84.0 | 79.3 | 60.4 | 76.0 |
| 15 | muse-spark-1.3 | 74.9 | 40.4 | 89.7 | 90.2 | 81.0 | 79.6 | 68.6 | 74.8 |
How this table is built
This table is rebuilt automatically, several times a day, by the same pipeline that refreshes the News section, using LiveBench's published per-task data. Category scores are the mean of each category's task scores, and the global average is the mean of the categories, matching how LiveBench presents its own leaderboard. If the upstream data is unreachable, the most recent successful snapshot stays in place and the date above tells you how fresh it is.
Content last reviewed October 12, 2026. Review dates are maintained as described on the About page.