Why this benchmark is useful
Many classic tests are memorized, saturated, or stale. LiveBench earns a slot because it explicitly refreshes the question pool on a schedule.
LiveBench
A contamination-limited benchmark that refreshes questions over time across reasoning, coding, math, data analysis, language, and instruction-following tasks.
What this does not measure
- Static benchmark mastery — the point is freshness, so old snapshots should not be treated like permanent model truth.
- Domain-specific reliability — a broad score can hide weak behavior in the domain your reader actually cares about.
- Cost, latency, or privacy — LiveBench is capability evidence, not product-readiness evidence.
Coverage map
- Task family
- Reasoning
- Format
- Freshly refreshed tasks across multiple categories; recent releases are published on the LiveBench site.
- Scoring
- Category and global averages by LiveBench release. Compare only inside the same release.
- Maintainer
- LiveBench team
How to read the scores
Always read the LiveBench release date with the score. A model's position on an older snapshot may not survive the next refresh.
What it does not cover
- Static benchmark mastery — the point is freshness, so old snapshots should not be treated like permanent model truth.
- Domain-specific reliability — a broad score can hide weak behavior in the domain your reader actually cares about.
- Cost, latency, or privacy — LiveBench is capability evidence, not product-readiness evidence.
Scores
21 rowsEvidence ledger
2121 rows
71.4%best score
0source-checked
1sources
2026-04-15source date
Where the rows land
0255075100
Normalized to this benchmark's axis (0–100). Open the table for raw units.
Newest receipts
- Claude Opus 4.671.4% · LiveBench leaderboard
- Claude Opus 4.870.8% · LiveBench leaderboard
- GPT 5.569.8% · LiveBench leaderboard
- o368.5% · LiveBench leaderboard
- Claude Sonnet 4.667.2% · LiveBench leaderboard
Scores use this benchmark's own unit and axis, not a universal quality score.
| # | Model | Release date | Score | Provenance | Trust |
|---|---|---|---|---|---|
| 1 | Claude Opus 4.6AnthropicLatest LiveBench release. Compare within same release only. | 2026-02-05 | 71.4% | LiveBench leaderboard2026-04-15 | Needs audit |
| 2 | Claude Opus 4.8Anthropic | 2026-05-28 | 70.8% | LiveBench leaderboard2026-04-15 | Needs audit |
| 3 | GPT 5.5OpenAI | 2026-04-23 | 69.8% | LiveBench leaderboard2026-04-15 | Needs audit |
| 4 | o3OpenAIStrong in math and reasoning categories; softer in instruction-following. | 2025-04-16 | 68.5% | LiveBench leaderboard2026-04-15 | Needs audit |
| 5 | Claude Sonnet 4.6Anthropic | 2026-02-17 | 67.2% | LiveBench leaderboard2026-04-15 | Needs audit |
| 6 | GPT 5.4OpenAI | 2026-03-05 | 66.4% | LiveBench leaderboard2026-04-15 | Needs audit |
| 7 | Gemini 3.1 ProGoogle | 2026-02-19 | 65.8% | LiveBench leaderboard2026-04-15 | Needs audit |
| 8 | Gemini 3.5 FlashGoogle | 2026-05-19 | 64.2% | LiveBench leaderboard2026-04-15 | Needs audit |
| 9 | DeepSeek R1DeepSeek | 2025-01-20 | 64.1% | LiveBench leaderboard2026-04-15 | Needs audit |
| 10 | Gemini 3 Flash PreviewGoogle | 2025-12-17 | 63.5% | LiveBench leaderboard2026-04-15 | Needs audit |
| 11 | Mistral Large 3Mistral | 2025-12-02 | 62.8% | LiveBench leaderboard2026-04-15 | Needs audit |
| 12 | Kimi K2.6Moonshot | 2026-04-20 | 62.4% | LiveBench leaderboard2026-04-15 | Needs audit |
| 13 | Llama 4 MaverickMeta | 2026-04-05 | 62.1% | LiveBench leaderboard2026-04-15 | Needs audit |
| 14 | DeepSeek V4DeepSeek | 2026-04-24 | 61.9% | LiveBench leaderboard2026-04-15 | Needs audit |
| 15 | Qwen 3.7 MaxAlibaba | 2026-05-20 | 61.6% | LiveBench leaderboard2026-04-15 | Needs audit |
| 16 | Kimi K2.5Moonshot | 2026-01-27 | 61.2% | LiveBench leaderboard2026-04-15 | Needs audit |
| 17 | Grok 4.3xAI | 2026-04-30 | 60.8% | LiveBench leaderboard2026-04-15 | Needs audit |
| 19 | MiniMax M3MiniMax | 2026-05-31 | 59.6% | LiveBench leaderboard2026-04-15 | Needs audit |
| 20 | Llama 4 ScoutMeta | 2026-04-05 | 58.9% | LiveBench leaderboard2026-04-15 | Needs audit |
| 21 | Claude Haiku 4.5Anthropic | 2025-10-01 | 57.8% | LiveBench leaderboard2026-04-15 | Needs audit |
| 22 | Grok 4.20 ReasoningxAI | 2026-03-05 | 56.4% | LiveBench leaderboard2026-04-15 | Needs audit |
What it covers
Primary source is livebench.ai. VerdictPal should record the LiveBench release label with every score row.
Where it breaks down
The broad average is a map, not a verdict. Drill into task families before using it to choose a tool.
Receipts
Sources and further reading
- LiveBench leaderboard· checked 2026-06-03
- LiveBench paper· checked 2026-06-03
- LiveBench
Questions about this benchmark
Every answer below is assembled from the dated fields on this page. Nothing is written separately for search.
What does LiveBench measure?
A contamination-limited benchmark that refreshes questions over time across reasoning, coding, math, data analysis, language, and instruction-following tasks.
What does a high LiveBench score not prove?
A strong LiveBench result says nothing about:
- Static benchmark mastery — the point is freshness, so old snapshots should not be treated like permanent model truth.
- Domain-specific reliability — a broad score can hide weak behavior in the domain your reader actually cares about.
- Cost, latency, or privacy — LiveBench is capability evidence, not product-readiness evidence.
How is LiveBench scored?
Category and global averages by LiveBench release. Compare only inside the same release. Task format: Freshly refreshed tasks across multiple categories; recent releases are published on the LiveBench site.
Is LiveBench saturated?
LiveBench is currently marked Active in the atlas.
Can LiveBench results be contaminated by training data?
Contamination risk for LiveBench is graded Low contamination. Treat every row on this page as a public claim with a source and a date, not as a controlled experiment.
The Pack · Editorial newsletter
New tools in your inbox. Free.
One short email when a tool ships or changes status. No tracking, no third-party analytics. Unsubscribe in one click.