AI leaderboard
Quale modello è in testa, in generale e per categoria. Questi sono i benchmark pubblici da cui sono costruiti i grafici che vedi su YouTube e X. Non pubblichiamo punteggi nostri: ogni numero appartiene al benchmark che lo ha prodotto, rilevato in una data e ricondotto alla fonte.
Real people are shown the same prompt answered by two anonymous models side by side and vote for the better answer. Millions of these blind head-to-head votes are fed into a Bradley-Terry statistical model (the successor to the Elo system it started with) which converts win/loss pairs into a single rating per model. A higher rating means people picked that model more often against strong opposition. This snapshot uses the 'style control' variant, which is the site's default: it statistically adjusts for answer length and formatting so a model cannot climb simply by writing longer, prettier replies.
Limitazioni: It measures which answer people LIKE, not which answer is CORRECT - a confident, well-written wrong answer can still win a vote. Voters are self-selected volunteers rather than a representative sample, prompts skew toward what that crowd chooses to type, and models with few votes have wide confidence intervals (ci_low/ci_high) that often overlap the models ranked above and below them. Treat small rank gaps as ties.
Fonte: LMArena (formerly LMSYS Chatbot Arena) ↗ · pubblicato 2026-08-11 · dati sottostanti ↗ · Dataset released under Creative Commons Attribution 4.0 (CC BY 4.0). Reuse is permitted with attribution - credit LMArena and link to the leaderboard.
| # | Modello | Organizzazione | Arena score (human preference) |
|---|---|---|---|
| 1 | claude-fable-5 | Anthropic | 1506 95% CI 1501 to 1512 21,304 votes |
| 2 | claude-opus-4-6-thinking | Anthropic | 1505 95% CI 1501 to 1508 72,425 votes |
| 3 | claude-opus-4-7-thinking | Anthropic | 1502 95% CI 1498 to 1506 60,222 votes |
| 4 | muse-spark-1.2 (xHigh) | Meta | 1498 95% CI 1488 to 1509 3,278 votes |
| 5 | claude-opus-4-6 | Anthropic | 1498 95% CI 1494 to 1501 76,386 votes |
| 6 | claude-opus-5-high | Anthropic | 1494 95% CI 1489 to 1499 19,498 votes |
| 7 | claude-opus-4-7 | Anthropic | 1494 95% CI 1490 to 1498 61,308 votes |
| 8 | claude-opus-5-max | Anthropic | 1490 95% CI 1483 to 1497 9,419 votes |
| 9 | qwen3.8-max | Alibaba | 1490 95% CI 1482 to 1498 6,789 votes |
| 10 | muse-spark-1.1 | Meta | 1489 95% CI 1483 to 1494 16,648 votes |
| 11 | muse-spark | Meta | 1488 95% CI 1482 to 1494 13,600 votes |
| 12 | kimi-k3-max | Moonshot AI | 1487 95% CI 1481 to 1493 11,762 votes |
| 13 | gemini-3.1-pro-preview | 1486 95% CI 1483 to 1490 94,814 votes | |
| 14 | gemini-3-pro | 1486 95% CI 1482 to 1489 41,509 votes | |
| 15 | gemini-3.6-flash | 1484 95% CI 1478 to 1490 13,559 votes | |
| 16 | gpt-5.5-high | OpenAI | 1482 95% CI 1477 to 1486 55,210 votes |
| 17 | claude-opus-4-8-thinking | Anthropic | 1481 95% CI 1477 to 1486 40,409 votes |
| 18 | gpt-5.6-sol-xhigh | OpenAI | 1481 95% CI 1475 to 1487 15,304 votes |
| 19 | gemini-3.5-flash-high | 1477 95% CI 1473 to 1482 25,613 votes | |
| 20 | gpt-5.5 | OpenAI | 1477 95% CI 1473 to 1481 56,513 votes |
Le barre sono scalate sull'intervallo visibile così le piccole differenze restano leggibili. Non partono da zero. Dove gli intervalli di confidenza si sovrappongono, i modelli sono statisticamente equivalenti: leggi il gruppo in cima come un gruppo, non come un ordine preciso.
Come leggere questi risultati e perché non concordano
LMArena (formerly LMSYS Chatbot Arena)
Real people are shown the same prompt answered by two anonymous models side by side and vote for the better answer. Millions of these blind head-to-head votes are fed into a Bradley-Terry statistical model (the successor to the Elo system it started with) which converts win/loss pairs into a single rating per model. A higher rating means people picked that model more often against strong opposition. This snapshot uses the 'style control' variant, which is the site's default: it statistically adjusts for answer length and formatting so a model cannot climb simply by writing longer, prettier replies.
Attenzione a: It measures which answer people LIKE, not which answer is CORRECT - a confident, well-written wrong answer can still win a vote. Voters are self-selected volunteers rather than a representative sample, prompts skew toward what that crowd chooses to type, and models with few votes have wide confidence intervals (ci_low/ci_high) that often overlap the models ranked above and below them. Treat small rank gaps as ties.
Pubblicato 2026-08-11 · classifica in tempo reale ↗ · Data: LMArena leaderboard dataset (CC BY 4.0).
LiveBench
A fixed set of test questions with objectively verifiable answers is run against each model and scored automatically against ground truth - no human voting and no AI judge, so the score is repeatable. This release spans 23 tasks grouped into 7 categories. Each category score is the average of its tasks, and the headline 'global average' is the average of the 7 category scores, so every category counts equally regardless of how many tasks it contains. Scores are percentages: 100 is perfect.
Attenzione a: Contamination-LIMITED, not contamination-proof: questions are refreshed from recent sources to reduce the chance a model simply memorised them during training, but that cannot be guaranteed. Scores reflect only these 23 tasks - they say nothing about tone, safety, speed or cost. Many entries are effort/thinking variants of the same underlying model (model_id shows the exact configuration tested), and a variant given more reasoning budget will usually outscore the cheaper default that most people actually use.
Pubblicato · classifica in tempo reale ↗ · Data: LiveBench 2026-06-25 release, livebench.ai.
Due unità diverse, mai un unico grafico
Una valutazione basata sulle preferenze umane e un punteggio di risposte corrette non possono condividere lo stesso asse, quindi questa pagina non li mescola mai in una stessa tabella. Un modello può primeggiare nell'uno e non nell'altro, ed è un segnale reale su ciò che sa fare, non una contraddizione: uno chiede "quale risposta hanno preferito le persone?", l'altro chiede "quale risposta era corretta?".
Cosa questa pagina non farà
Non ti dirà quale AI usare. Chi guida i benchmark spesso non è lo strumento giusto per il tuo lavoro, prezzo, disponibilità, integrazioni, lunghezza del contesto e quanto segue BENE LE TUE istruzioni contano di solito più di un punto o due di punteggio. Quale AI dovrei usare? →