{"categories":["LLM"],"content":{"markdown":"Ein Benchmark-Wert sagt für sich wenig. Aussagekräftig wird er, wenn man sieht, wie schnell er sich bewegt hat: welches Modell wann den Rekord setzte — und wie lange der Test hielt, bevor der nächste erfunden werden musste.\n\nDiese Seite führt dieses Protokoll für 16 Benchmarks, von der GPT-2-Zeit bis heute. Zu jedem sehen Sie jedes Modell, das bei seinem Erscheinen alle früheren übertraf, und die Spitze von heute. Die Zahlen stammen aus dem offenen Benchmark-Archiv von Epoch AI — 2.621 Ergebnisse in den 16 hier gezeigten Benchmarks. Wir haben kein Modell selbst getestet und keinen Wert verändert; unser Anteil sind die Auswahl, die Rekordlinien und der Vergleich über die Zeit.\n\n## So lesen Sie die Tabellen\n\n**Bestmarke** heißt: bester Wert unter allen Modellen, die bis zu diesem Datum erschienen sind. Das Datum ist das Erscheinungsdatum des Modells, wie das Archiv es führt, nicht der Tag des Tests. Ein später getestetes Modell steht trotzdem an seinem Erscheinungsdatum.\n\nWerte sind zwischen Benchmarks nicht vergleichbar, und auch innerhalb nicht immer: Ergebnisse, die Epoch AI selbst erhoben hat, folgen einem festen Aufbau; übernommene Ergebnisse stammen aus Fachartikeln und Ranglisten mit eigenem Aufbau. Bei jedem Benchmark steht, woher seine Werte kommen.\n\nSteht bei einem Benchmark „keine neuen Ergebnisse seit“, ist das meist ein Zeichen von Erfolg, nicht von Vernachlässigung: Die besten Modelle liegen nahe am Maximum, und das Feld ist zu härteren Tests weitergezogen.\n\n## Die Bestmarken\n\n{{\u003c llm-benchmarks \u003e}}\n\n## Was die Rekordlinien zeigen\n\n**Tests sind immer schneller verbraucht.** Auf MMLU stand der Rekord bei 86,4 %, als GPT-4 im März 2023 erschien, und hat sich seitdem kaum bewegt — der Test war ausgereizt. ARC-AGI-2 wurde im März 2025 als Rätselsatz eingeführt, den Modelle nicht lösen konnten; das erste Modell über zehn Prozent folgte im Juli 2025, und der Rekord steht heute bei 95,0 %.\n\n**Die Länge der Aufgaben wächst, nicht nur der Punktwert.** METRs Zeithorizont ist hier das einzige Maß ohne Deckel bei 100. GPT-4 bewältigte 2023 Aufgaben von etwa 5 Minuten; der aktuelle Rekord liegt bei über 17 Stunden.\n\n**Eine Zahl verbirgt den Aufbau.** Dasselbe Modell taucht mit verschiedenen Werten auf — je nach Denkaufwand oder Agenten-Software drumherum; die Namen in den Tabellen tragen diese Einstellung in Klammern. Wer zwei Modelle für das eigene Projekt vergleicht, vergleicht sie in der Einstellung, die tatsächlich zum Einsatz kommt.\n\nWer mit diesen Modellen baut, für den sind Benchmarks der Anfang der Auswahl, nicht ihr Ende. Das [LLM-Ranking](/de/llm/ranking/) stellt alle Modelle auf eine Skala, und in der [LLM-Datenbank](/de/llm/) lassen sie sich filtern und je vier nebeneinander vergleichen.","text":"Ein Benchmark-Wert sagt für sich wenig. Aussagekräftig wird er, wenn man sieht, wie schnell er sich bewegt hat: welches Modell wann den Rekord setzte — und wie lange der Test hielt, bevor der nächste erfunden werden musste.\nDiese Seite führt dieses Protokoll für 16 Benchmarks, von der GPT-2-Zeit bis heute. Zu jedem sehen Sie jedes Modell, das bei seinem Erscheinen alle früheren übertraf, und die Spitze von heute. Die Zahlen stammen aus dem offenen Benchmark-Archiv von Epoch AI — 2.621 Ergebnisse in den 16 hier gezeigten Benchmarks. Wir haben kein Modell selbst getestet und keinen Wert verändert; unser Anteil sind die Auswahl, die Rekordlinien und der Vergleich über die Zeit.\nSo lesen Sie die Tabellen Bestmarke heißt: bester Wert unter allen Modellen, die bis zu diesem Datum erschienen sind. Das Datum ist das Erscheinungsdatum des Modells, wie das Archiv es führt, nicht der Tag des Tests. Ein später getestetes Modell steht trotzdem an seinem Erscheinungsdatum.\nWerte sind zwischen Benchmarks nicht vergleichbar, und auch innerhalb nicht immer: Ergebnisse, die Epoch AI selbst erhoben hat, folgen einem festen Aufbau; übernommene Ergebnisse stammen aus Fachartikeln und Ranglisten mit eigenem Aufbau. Bei jedem Benchmark steht, woher seine Werte kommen.\nSteht bei einem Benchmark „keine neuen Ergebnisse seit“, ist das meist ein Zeichen von Erfolg, nicht von Vernachlässigung: Die besten Modelle liegen nahe am Maximum, und das Feld ist zu härteren Tests weitergezogen.\nDie Bestmarken Alle Benchmarks im Überblick BenchmarkModelle von–bisBestmarke heuteModell Epoch Capabilities Index2023–2026167,3Claude Opus 5.5 Anthropic METR Time Horizon2019–202617 Std. 25 minClaude Mythos Preview (Early) Anthropic MMLU2021–202488,1 %GPT-4o OpenAI GPQA Diamond2023–202695,8 %GPT-6 Astra (max) OpenAI Humanity\u0026#39;s Last Exam2024–202654,8 %GPT 6 Astra OpenAI SimpleQA Verified2024–202675,6 %GPT-6 Astra (max) OpenAI GSM8K2020–202494,5 %DS-Coder-V2-Instruct DeepSeek MATH Level 52023–202598,1 %GPT-5 (high) OpenAI FrontierMath (Tiers 1–3)2024–202693,7 %GPT-6 Astra (max) OpenAI SWE-bench Verified2024–202683,5 %Claude Opus 4.7 (max) Anthropic Aider Polyglot2024–202588,0 %GPT-5 (high) OpenAI Terminal-Bench2025–202684,7 %GPT-5.5 OpenAI ARC-AGI2024–202698,5 %Claude Fable 5 (XHigh) Anthropic ARC-AGI-22024–202695,0 %GPT-6 Astra (Max) OpenAI HellaSwag2019–202495,3 %GPT-4 OpenAI WinoGrande2019–202489,2 %LLaMA-3.1 405B Base Meta AI Gesamtleistung Epoch Capabilities Index veröffentlicht 2025 · Modelle von 2023 bis 2026 · 274 Modelle im Archiv · läuft · Werte: Epoch AI\nFasst die Ergebnisse von mehr als 50 Benchmarks zu einer Zahl je Modell zusammen, sodass sich Modelle vergleichen lassen, auch wenn sie nie auf denselben Benchmarks geprüft wurden. Ein Index, kein Prozentwert: Epoch AI skaliert ihn so, dass Claude 3.5 Sonnet bei 130 und GPT-5 bei 150 steht.\nBestmarken nach Erscheinungsdatum des Modells DatumModellWert 24.02.2023LLaMA-65B Meta AI110,2 14.03.2023GPT-4 (Mar 2023) OpenAI125,9 25.01.2024GPT-4 Turbo (Nov 2023) OpenAI126,5 29.02.2024Claude 3 Opus Anthropic126,9 09.04.2024GPT-4 Turbo (Apr 2024) OpenAI127,2 13.05.2024GPT-4o (May 2024) OpenAI129,0 20.06.2024Claude 3.5 Sonnet Anthropic130,0 12.09.2024o1-mini OpenAI135,8 17.12.2024o1 OpenAI141,9 31.03.2025Gemini 2.5 Pro (Mar 2025) Google DeepMind144,2 16.04.2025o3 OpenAI146,9 10.06.2025o3-pro OpenAI147,4 07.08.2025GPT-5 OpenAI150,0 07.10.2025GPT-5 Pro OpenAI150,3 18.11.2025Gemini 3 Pro Google DeepMind152,9 11.12.2025GPT-5.2 Pro OpenAI155,4 05.02.2026GPT-5.3 Codex OpenAI156,8 05.03.2026GPT-5.4 Pro OpenAI158,9 23.04.2026GPT-5.5 Pro OpenAI162,1 24.07.2026Claude Opus 5 Anthropic162,8 01.09.2026Claude Fable 5.1 Anthropic164,7 03.09.2026GPT-6 Astra OpenAI166,4 22.09.2026Claude Opus 5.5 Anthropic167,3 Spitze heute: Claude Opus 5.5 (167,3) · GPT-6 Astra (166,4) · GPT-6.1 Sol (166,1) · Claude Sonnet 5.5 (165,0) · Claude Fable 5.1 (164,7)\nMETR Time Horizon veröffentlicht 2025 · Modelle von 2019 bis 2026 · 39 Modelle im Archiv · läuft · Werte: METR\nDie Länge der Aufgaben, die ein Modell in etwa der Hälfte der Fälle richtig löst — gemessen daran, wie lange ein menschlicher Experte dafür braucht. Die Zahl ist eine Dauer, kein Punktwert: Ein längerer Horizont heißt, das Modell kann länger selbstständig arbeiten.\nBestmarken nach Erscheinungsdatum des Modells DatumModellWert 05.11.2019GPT-2 (1.5B) OpenAI3 Sek. 14.03.2023GPT-4 (Mar 2023) OpenAI5,36 min 25.01.2024GPT-4 Turbo Preview (January 2024) OpenAI5,37 min 09.04.2024GPT-4 Turbo (Apr 2024) OpenAI6,57 min 13.05.2024GPT-4o (May 2024) OpenAI9,17 min 20.06.2024Claude 3.5 Sonnet (Jun 2024) Anthropic11 min 12.09.2024o1-preview OpenAI20 min 22.10.2024Claude 3.5 Sonnet (Oct 2024) Anthropic21 min 17.12.2024o1 (medium) OpenAI39 min 24.02.2025Claude 3.7 Sonnet Anthropic1 Std. 16.04.2025o3 (unknown thinking) OpenAI2 Std. 07.08.2025GPT-5 (high) OpenAI3 Std. 23 min 18.11.2025Gemini 3 Pro Preview Google DeepMind3 Std. 44 min 24.11.2025Claude Opus 4.5 (no thinking) Anthropic4 Std. 53 min 11.12.2025GPT-5.2 (high) OpenAI5 Std. 52 min 05.02.2026Claude Opus 4.6 (unknown thinking) Anthropic11 Std. 59 min 07.04.2026Claude Mythos Preview (Early) Anthropic17 Std. 25 min Die Balken sind logarithmisch skaliert: Jeder Schritt nach rechts ist eine Vervielfachung.\nSpitze heute: Claude Mythos Preview (Early) (17 Std. 25 min) · Claude Opus 4.6 (unknown thinking) (11 Std. 59 min) · Gemini 3.1 Pro Preview (6 Std. 24 min) · GPT-5.2 (high) (5 Std. 52 min) · GPT-5.3 Codex (5 Std. 50 min)\nWissen und Schlussfolgern MMLU veröffentlicht 2020 · Modelle von 2021 bis 2024 · 158 Modelle im Archiv · keine neuen Ergebnisse seit 2024 · Werte: Fachartikel und Stanford HELM, gesammelt von Epoch AI\nMultiple-Choice-Fragen aus 57 Fächern, vom Schulniveau bis zur Berufsprüfung. Jahrelang der Standardtest für Allgemeinwissen; heute liegen die besten Modelle so nah an der Obergrenze, dass kaum noch neue Ergebnisse erscheinen.\nDie bekannten 43,9 % von GPT-3 führt das Archiv unter der API-Version vom Januar 2022; als Bestmarke tauchen sie deshalb hier nicht auf.\nBestmarken nach Erscheinungsdatum des Modells DatumModellWert 05.08.2021GPT-J-6B EleutherAI, LAION25,7 % 08.12.2021Gopher DeepMind60,0 % 15.03.2022GPT-3.5 OpenAI70,0 % 14.03.2023GPT-4 (initial release) OpenAI86,4 % 20.06.2024Claude 3.5 Sonnet (20240620) Anthropic86,5 % 24.09.2024Gemini 1.5 Pro (002) Google DeepMind86,9 % 22.10.2024Claude 3.5 Sonnet (20241022) Anthropic87,3 % 20.11.2024GPT-4o OpenAI88,1 % Spitze heute: GPT-4o (88,1 %) · Claude 3.5 Sonnet (20241022) (87,3 %) · DeepSeek v3 (87,2 %) · Gemini 1.5 Pro (002) (86,9 %) · Claude 3.5 Sonnet (20240620) (86,5 %)\nGPQA Diamond veröffentlicht 2023 · Modelle von 2023 bis 2026 · 214 Modelle im Archiv · läuft · Werte: Epoch AI (eigene Läufe)\n198 Multiple-Choice-Fragen aus Biologie, Physik und Chemie auf Graduiertenniveau, so gebaut, dass Fachfremde selbst mit Webzugang scheitern. Raten ergibt 25 %.\nBestmarken nach Erscheinungsdatum des Modells DatumModellWert 14.03.2023GPT-4 (Mar 2023) OpenAI35,7 % 06.11.2023GPT-4 Turbo Preview (Nov 2023) OpenAI42,4 % 29.02.2024Claude 3 Opus Anthropic47,2 % 13.05.2024GPT-4o (May 2024) OpenAI48,9 % 20.06.2024Claude 3.5 Sonnet (Jun 2024) Anthropic54,0 % 12.09.2024o1-mini (high) OpenAI62,4 % 17.12.2024o1 (high) OpenAI76,8 % 31.01.2025o3-mini (high) OpenAI77,0 % 24.02.2025Claude 3.7 Sonnet (64k thinking) Anthropic79,7 % 25.03.2025Gemini 2.5 Pro Exp (Mar 2025) Google DeepMind83,8 % 05.06.2025Gemini 2.5 Pro Preview (Jun 2025) Google DeepMind84,8 % 09.07.2025Grok 4 xAI87,0 % 13.11.2025GPT-5.1 (high) OpenAI87,6 % 18.11.2025Gemini 3 Pro Preview Google DeepMind92,6 % 19.02.2026Gemini 3.1 Pro Preview Google DeepMind94,4 % 05.03.2026GPT-5.4 Pro (xhigh) OpenAI94,6 % 13.08.2026Gemini 3.7 Flash (high) Google DeepMind94,8 % 02.09.2026Gemini 3.8 Flash (high) Google DeepMind95,4 % 03.09.2026GPT-6 Astra (max) OpenAI95,8 % Spitze heute: GPT-6 Astra (max) (95,8 %) · Claude Sonnet 5.5 (max) (95,6 %) · Gemini 3.8 Flash (high) (95,4 %) · GPT-6.1 Sol (max) (95,4 %) · Gemini 3.7 Flash (high) (94,8 %)\nHumanity\u0026#39;s Last Exam veröffentlicht 2025 · Modelle von 2024 bis 2026 · 44 Modelle im Archiv · läuft · Werte: Rangliste von Humanity\u0026#39;s Last Exam, gesammelt von Epoch AI\nVon Fachleuten geschriebene Fragen an der Grenze des menschlichen Wissens aus vielen Gebieten — gebaut, um schwer zu bleiben, nachdem ältere Tests gelöst waren.\nBestmarken nach Erscheinungsdatum des Modells DatumModellWert 24.09.2024Gemini-1.5-Pro-002 Google DeepMind4,6 % 17.12.2024o1 (December 2024) OpenAI8,0 % 19.03.2025o1 Pro OpenAI8,1 % 25.03.2025Gemini 2.5 Pro Experimental (March 2025) Google DeepMind18,2 % 16.04.2025o3 (high) (April 2025) OpenAI20,3 % 05.06.2025gemini-2.5-pro-preview-06-05 Google DeepMind21,6 % 07.08.2025gpt-5-2025-08-07 OpenAI25,3 % 07.10.2025gpt-5-pro-2025-10-06 OpenAI31,6 % 18.11.2025gemini-3-pro-preview Google DeepMind37,5 % 19.02.2026Gemini 3.1 Pro Preview (thinking high) Google DeepMind46,4 % 01.09.2026Fable 5.1 (xhigh) Anthropic46,5 % 03.09.2026GPT 6 Astra OpenAI54,8 % Spitze heute: GPT 6 Astra (54,8 %) · Fable 5.1 (xhigh) (46,5 %) · Gemini 3.1 Pro Preview (thinking high) (46,4 %) · Gemini 3.8 Flash (44,5 %) · GPT-5.4 Pro (44,3 %)\nSimpleQA Verified veröffentlicht 2025 · Modelle von 2024 bis 2026 · 79 Modelle im Archiv · läuft · Werte: Epoch AI (eigene Läufe)\nKurze Faktenfragen mit genau einer richtigen Antwort, beantwortet ohne Hilfsmittel. Zeigt, wie viel ein Modell wirklich weiß — und wie oft es eine Antwort erfindet, statt Nichtwissen zuzugeben.\nBestmarken nach Erscheinungsdatum des Modells DatumModellWert 29.02.2024Claude 3 Opus Anthropic12,6 % 06.08.2024GPT-4o (Aug 2024) OpenAI26,0 % 17.12.2024o1 (high) OpenAI41,1 % 16.04.2025o3 (high) OpenAI49,4 % 07.08.2025GPT-5 (high) OpenAI50,1 % 17.12.2025Gemini 3 Flash Preview Google DeepMind66,8 % 19.02.2026Gemini 3.1 Pro Preview Google DeepMind73,5 % 03.09.2026GPT-6 Astra (max) OpenAI75,6 % Spitze heute: GPT-6 Astra (max) (75,6 %) · GPT-6.1 Sol (max) (73,9 %) · Gemini 3.1 Pro Preview (73,5 %) · Claude Opus 5.5 (max) (72,2 %) · Claude Fable 5.1 (max) (70,8 %)\nMathematik GSM8K veröffentlicht 2021 · Modelle von 2020 bis 2024 · 146 Modelle im Archiv · keine neuen Ergebnisse seit 2024 · Werte: Fachartikel und Stanford HELM, gesammelt von Epoch AI\nTextaufgaben auf Grundschulniveau, die mehrere Rechenschritte brauchen. Einst ein harter Test für logisches Schließen, heute praktisch gelöst.\nBestmarken nach Erscheinungsdatum des Modells DatumModellWert 22.06.2020babbage (1.3B) OpenAI0,7 % 15.03.2022code-davinci-002 OpenAI56,8 % 28.11.2022GPT-3.5 (8-shot) OpenAI78,2 % 14.03.2023GPT-4 OpenAI92,0 % 17.06.2024DS-Coder-V2-Instruct DeepSeek94,5 % Spitze heute: DS-Coder-V2-Instruct (94,5 %) · Qwen2.5-Coder-14B-Instruct (94,2 %) · Qwen2.5-Coder-32B-Instruct (93,0 %) · GPT-4 (92,0 %) · GPT-4 (92,0 %)\nMATH Level 5 veröffentlicht 2021 · Modelle von 2023 bis 2025 · 94 Modelle im Archiv · läuft · Werte: Epoch AI (eigene Läufe)\nDie schwerste Stufe des MATH-Datensatzes: Wettbewerbsaufgaben aus Algebra, Geometrie, Zahlentheorie und Wahrscheinlichkeitsrechnung.\nBestmarken nach Erscheinungsdatum des Modells DatumModellWert 13.06.2023GPT-4 (Jun 2023) OpenAI23,0 % 06.11.2023GPT-4 Turbo Preview (Nov 2023) OpenAI40,0 % 09.04.2024GPT-4 Turbo (Apr 2024) OpenAI46,7 % 13.05.2024GPT-4o (May 2024) OpenAI51,0 % 20.06.2024Claude 3.5 Sonnet (Jun 2024) Anthropic51,7 % 18.07.2024GPT-4o mini OpenAI52,6 % 06.08.2024GPT-4o (Aug 2024) OpenAI53,3 % 12.09.2024o1-mini (high) OpenAI89,2 % 17.12.2024o1 (high) OpenAI94,7 % 31.01.2025o3-mini (high) OpenAI96,5 % 16.04.2025o4-mini (high) OpenAI97,8 % 07.08.2025GPT-5 (high) OpenAI98,1 % Spitze heute: GPT-5 (high) (98,1 %) · o4-mini (high) (97,8 %) · o3 (high) (97,8 %) · GPT-5 mini (high) (97,8 %) · Claude Sonnet 4.5 (32k thinking) (97,7 %)\nFrontierMath (Tiers 1–3) veröffentlicht 2026 · Modelle von 2024 bis 2026 · 83 Modelle im Archiv · läuft · Werte: Epoch AI (eigene Läufe)\nEigens geschriebene, unveröffentlichte Mathematikaufgaben von Mathematikern, vom fortgeschrittenen Studium bis zur Forschung. Die Aufgaben bleiben unter Verschluss, damit Modelle sie nicht aus dem Training kennen können.\nDas Archiv führt diese zweite Fassung des Aufgabensatzes seit Juni 2026, auch mit Ergebnissen älterer Modelle.\nBestmarken nach Erscheinungsdatum des Modells DatumModellWert 25.01.2024GPT-3.5 Turbo (Jan 2024) OpenAI0,0 % 09.04.2024GPT-4 Turbo (Apr 2024) OpenAI0,7 % 17.12.2024o1 (high) OpenAI14,7 % 31.01.2025o3-mini (high) OpenAI18,6 % 16.04.2025o4-mini (high) OpenAI36,1 % 07.08.2025GPT-5 (high) OpenAI55,4 % 07.10.2025GPT-5 Pro OpenAI55,8 % 11.12.2025GPT-5.2 Pro OpenAI74,0 % 05.03.2026GPT-5.4 Pro (xhigh) OpenAI82,5 % 23.04.2026GPT-5.5 Pro (xhigh) OpenAI87,7 % 09.07.2026GPT-5.6 Sol (max) OpenAI89,1 % 01.09.2026Claude Fable 5.1 (max) Anthropic90,2 % 03.09.2026GPT-6 Astra (max) OpenAI93,7 % Spitze heute: GPT-6 Astra (max) (93,7 %) · GPT-6.1 Sol (max) (93,7 %) · Claude Opus 5.5 (max) (91,2 %) · Claude Fable 5.1 (max) (90,2 %) · GPT-6 Sol (max) (89,8 %)\nProgrammieren und Agenten SWE-bench Verified veröffentlicht 2024 · Modelle von 2024 bis 2026 · 32 Modelle im Archiv · läuft · Werte: Epoch AI (eigene Läufe)\n500 echte Fehlerberichte aus Open-Source-Projekten in Python, von Menschen geprüft. Das Modell muss den Code so ändern, dass die Tests des Projekts bestehen.\nDies sind Epoch AIs eigene Läufe mit einem festen Aufbau; Anbieter melden mit eigenem Werkzeug oft höhere Werte.\nBestmarken nach Erscheinungsdatum des Modells DatumModellWert 20.11.2024GPT-4o (Nov 2024) OpenAI31,0 % 24.02.2025Claude 3.7 Sonnet Anthropic61,0 % 16.04.2025o3 (medium) OpenAI62,3 % 22.05.2025Claude Opus 4 Anthropic70,7 % 05.08.2025Claude Opus 4.1 Anthropic73,3 % 07.08.2025GPT-5 (high) OpenAI73,6 % 24.11.2025Claude Opus 4.5 (no thinking) Anthropic76,7 % 05.02.2026Claude Opus 4.6 (no thinking) Anthropic78,7 % 16.04.2026Claude Opus 4.7 (max) Anthropic83,5 % Spitze heute: Claude Opus 4.7 (max) (83,5 %) · GPT-5.5 (xhigh) (80,6 %) · Gemini 3.5 Flash (79,3 %) · Claude Opus 4.6 (no thinking) (78,7 %) · GLM-5.2 (78,7 %)\nAider Polyglot veröffentlicht 2024 · Modelle von 2024 bis 2025 · 54 Modelle im Archiv · läuft · Werte: Aider-Rangliste, gesammelt von Epoch AI\n225 schwierige Programmieraufgaben in C\u0026#43;\u0026#43;, Go, Java, JavaScript, Python und Rust. Misst, ob ein Modell vorhandenen Code korrekt ändern kann — mit einem zweiten Versuch nach Sicht der Testausgabe.\nBestmarken nach Erscheinungsdatum des Modells DatumModellWert 18.07.2024GPT-4o mini OpenAI3,6 % 06.08.2024GPT-4o (Aug 2024) OpenAI23,1 % 12.09.2024o1-mini (unknown thinking) OpenAI32,9 % 22.10.2024Claude 3.5 Sonnet (Oct 2024) Anthropic51,6 % 17.12.2024o1 (high) OpenAI61,7 % 24.02.2025Claude 3.7 Sonnet (32k thinking) Anthropic64,9 % 25.03.2025Gemini 2.5 Pro Exp (Mar 2025) Google DeepMind72,9 % 16.04.2025o3 (high) OpenAI81,3 % 05.06.2025Gemini 2.5 Pro Preview (Jun 2025) Google DeepMind83,1 % 10.06.2025o3-pro OpenAI84,9 % 07.08.2025GPT-5 (high) OpenAI88,0 % Spitze heute: GPT-5 (high) (88,0 %) · o3-pro (84,9 %) · Gemini 2.5 Pro Preview (Jun 2025) (83,1 %) · o3 (high) (81,3 %) · Grok 4 (79,6 %)\nTerminal-Bench veröffentlicht 2025 · Modelle von 2025 bis 2026 · 45 Modelle im Archiv · läuft · Werte: Terminal-Bench-Rangliste, gesammelt von Epoch AI\nAufgaben, die ein Agent in einer echten Kommandozeile lösen muss: Software einrichten, ein System reparieren, Daten verarbeiten. Der Wert gehört zu einem Modell samt Agenten-Software; wir zeigen das beste Ergebnis je Modell.\nBestmarken nach Erscheinungsdatum des Modells DatumModellWert 17.06.2025Gemini 2.5 Flash Google DeepMind17,1 % 09.07.2025Grok 4 xAI27,2 % 12.07.2025Kimi K2 Instruct Moonshot27,8 % 05.08.2025Claude Opus 4.1 Anthropic38,0 % 07.08.2025GPT-5 OpenAI49,6 % 12.11.2025GPT-5.1-Codex-Mini OpenAI61,6 % 18.11.2025Gemini 3 Pro Google DeepMind69,4 % 05.02.2026Claude Opus 4.6 Anthropic79,8 % 19.02.2026Gemini 3.1 Pro Google DeepMind80,2 % 05.03.2026GPT-5.4 OpenAI81,8 % 23.04.2026GPT-5.5 OpenAI84,7 % Spitze heute: GPT-5.5 (84,7 %) · GPT-5.4 (81,8 %) · Gemini 3.1 Pro (80,2 %) · Claude Opus 4.7 (80,2 %) · Claude Opus 4.6 (79,8 %)\nAbstraktion ARC-AGI veröffentlicht 2019 · Modelle von 2024 bis 2026 · 91 Modelle im Archiv · läuft · Werte: ARC-Prize-Rangliste, gesammelt von Epoch AI\nGitterrätsel, bei denen aus wenigen Beispielen eine Regel abgeleitet und auf einen neuen Fall angewandt werden muss. Für Menschen leicht, für Sprachmodelle lange unerreichbar — der Test, ob etwas Neues gelernt statt nur abgerufen wird.\nBestmarken nach Erscheinungsdatum des Modells DatumModellWert 12.09.2024o1-preview OpenAI18,0 % 17.12.2024o1 (Medium) OpenAI30,7 % 31.01.2025o3-mini (High) OpenAI34,5 % 16.04.2025o3 (High) OpenAI60,8 % 09.07.2025Grok 4 (Thinking) xAI66,7 % 07.10.2025GPT-5 Pro OpenAI70,2 % 13.11.2025GPT-5.1 (Thinking, High) OpenAI72,8 % 18.11.2025Gemini 3 Pro Google DeepMind75,0 % 24.11.2025Claude Opus 4.5 (64k thinking) Anthropic80,0 % 11.12.2025GPT-5.2 Pro (X-High) OpenAI90,5 % 05.02.2026Claude Opus 4.6 (120K, High) Anthropic94,0 % 12.02.2026Gemini 3 Deep Think (2/26) Google DeepMind96,0 % 19.02.2026Gemini 3.1 Pro (Preview) Google DeepMind98,0 % 09.06.2026Claude Fable 5 (XHigh) Anthropic98,5 % Spitze heute: Claude Fable 5 (XHigh) (98,5 %) · Gemini 3.8 Flash (High) (98,5 %) · GPT-6 Astra (XHigh) (98,5 %) · Claude Opus 5.5 (High) (98,5 %) · GPT-6.1 Sol (XHigh) (98,5 %)\nARC-AGI-2 veröffentlicht 2025 · Modelle von 2024 bis 2026 · 86 Modelle im Archiv · läuft · Werte: ARC-Prize-Rangliste, gesammelt von Epoch AI\nDer Nachfolger von ARC-AGI aus dem Jahr 2025, mit Rätseln, die gezielt den Methoden widerstehen sollen, mit denen die erste Fassung gelöst wurde.\nBestmarken nach Erscheinungsdatum des Modells DatumModellWert 18.07.2024GPT-4o-mini OpenAI0,0 % 12.09.2024o1-mini OpenAI0,8 % 20.01.2025Deepseek R1 DeepSeek1,3 % 31.01.2025o3-mini (High) OpenAI3,0 % 16.04.2025o3 (High) OpenAI6,5 % 22.05.2025Claude Opus 4 (Thinking 16K) Anthropic8,6 % 09.07.2025Grok 4 (Thinking) xAI16,0 % 07.10.2025GPT-5 Pro OpenAI18,3 % 18.11.2025Gemini 3 Pro Google DeepMind31,1 % 24.11.2025Opus 4.5 (Thinking, 64K) Anthropic37,6 % 11.12.2025GPT-5.2 Pro (High) OpenAI54,2 % 05.02.2026Claude Opus 4.6 (120K, High) Anthropic69,2 % 12.02.2026Gemini 3 Deep Think (2/26) Google DeepMind84,6 % 23.04.2026GPT-5.5 (xHigh) OpenAI85,0 % 09.06.2026Claude Fable 5 (Max) Anthropic89,2 % 09.07.2026GPT-5.6 Sol (Max) OpenAI92,5 % 03.09.2026GPT-6 Astra (Max) OpenAI95,0 % Spitze heute: GPT-6 Astra (Max) (95,0 %) · GPT-6.1 Sol (Max) (94,2 %) · Claude Opus 5.5 (High) (93,3 %) · GPT-5.6 Sol (Max) (92,5 %) · Claude Opus 5 (Max) (90,4 %)\nSprachverständnis (Klassiker) HellaSwag veröffentlicht 2019 · Modelle von 2019 bis 2024 · 97 Modelle im Archiv · keine neuen Ergebnisse seit 2024 · Werte: Fachartikel und Stanford HELM, gesammelt von Epoch AI\nAus vier Möglichkeiten die plausibelste Fortsetzung einer Alltagsszene wählen. Ein Test für gesunden Menschenverstand von 2019, den Menschen trivial finden; Raten ergibt 25 %.\nBestmarken nach Erscheinungsdatum des Modells DatumModellWert 05.11.2019GPT2-XL-1.5B OpenAI40,0 % 22.06.2020babbage (1.3B) OpenAI55,5 % 05.08.2021GPT-J-6B EleutherAI, LAION66,2 % 11.10.2021Megatron-NLG (530B) Microsoft, Nvidia82,4 % 04.04.2022PaLM-540B Google Research83,8 % 28.11.2022GPT-3.5 OpenAI85,5 % 14.03.2023GPT-4 OpenAI95,3 % Spitze heute: GPT-4 (95,3 %) · LLaMA-3.1 405B Base (89,2 %) · Falcon-180B (89,0 %) · DeepSeek-V3 Base (88,9 %) · DeepSeek-V2 Base (87,1 %)\nWinoGrande veröffentlicht 2019 · Modelle von 2019 bis 2024 · 84 Modelle im Archiv · keine neuen Ergebnisse seit 2024 · Werte: Fachartikel, gesammelt von Epoch AI\nEntscheiden, worauf sich ein mehrdeutiges Wort im Satz bezieht — lösbar nur mit Verständnis der Situation. Raten ergibt 50 %.\nBestmarken nach Erscheinungsdatum des Modells DatumModellWert 05.11.2019GPT-2-XL 1.5B OpenAI58,3 % 05.08.2021GPT-J-6B EleutherAI, LAION64,5 % 11.10.2021Megatron-NLG (530B) Microsoft, Nvidia78,9 % 13.12.2021GLaM Google79,2 % 15.03.2022GPT-3.5 OpenAI81,6 % 04.04.20225-shot PaLM Google Research85,1 % 14.03.2023GPT-4 OpenAI87,5 % 29.02.2024Claude 3 Opus Anthropic88,5 % 23.07.2024LLaMA-3.1 405B Base Meta AI89,2 % Spitze heute: LLaMA-3.1 405B Base (89,2 %) · Claude 3 Opus (88,5 %) · GPT-4 (87,5 %) · Falcon-180B (87,1 %) · UNICORN (86,6 %)\nQuelle und Lizenz Daten: Epoch AI, „Capabilities \u0026amp; benchmarking“, abgerufen am 12.10.2026, Lizenz CC BY 4.0. Ergebnisse, die Epoch AI von anderen übernommen hat, behalten ihre ursprüngliche Quelle; sie ist bei jedem Benchmark genannt. ABXK hat daraus die Bestmarken und die Spitzengruppen berechnet und keinen Wert verändert.\nWas die Rekordlinien zeigen Tests sind immer schneller verbraucht. Auf MMLU stand der Rekord bei 86,4 %, als GPT-4 im März 2023 erschien, und hat sich seitdem kaum bewegt — der Test war ausgereizt. ARC-AGI-2 wurde im März 2025 als Rätselsatz eingeführt, den Modelle nicht lösen konnten; das erste Modell über zehn Prozent folgte im Juli 2025, und der Rekord steht heute bei 95,0 %.\nDie Länge der Aufgaben wächst, nicht nur der Punktwert. METRs Zeithorizont ist hier das einzige Maß ohne Deckel bei 100. GPT-4 bewältigte 2023 Aufgaben von etwa 5 Minuten; der aktuelle Rekord liegt bei über 17 Stunden.\nEine Zahl verbirgt den Aufbau. Dasselbe Modell taucht mit verschiedenen Werten auf — je nach Denkaufwand oder Agenten-Software drumherum; die Namen in den Tabellen tragen diese Einstellung in Klammern. Wer zwei Modelle für das eigene Projekt vergleicht, vergleicht sie in der Einstellung, die tatsächlich zum Einsatz kommt.\nWer mit diesen Modellen baut, für den sind Benchmarks der Anfang der Auswahl, nicht ihr Ende. Das LLM-Ranking stellt alle Modelle auf eine Skala, und in der LLM-Datenbank lassen sie sich filtern und je vier nebeneinander vergleichen.\n","words":3012},"date":"2026-10-12","description":"Welches Modell hielt wann den Rekord? 16 LLM-Benchmarks von GPT-2 bis heute, mit jeder Bestmarke nach Erscheinungsdatum und der Spitze von heute — gebaut aus dem offenen Benchmark-Archiv von Epoch AI.","image":"https://abxk.ai/images/cover/llm-benchmarks.webp","json":"https://abxk.ai/de/llm/benchmarks/index.json","kind":"page","language":"de","llm_benchmarks":{"benchmarks":[{"active":true,"best":{"date":"2026-09-22","model":"Claude Opus 5.5","org":"Anthropic","src":"","value":167.3},"first_model":"2023-02-24","frontier":[{"date":"2023-02-24","model":"LLaMA-65B","org":"Meta AI","src":"","value":110.2},{"date":"2023-03-14","model":"GPT-4 (Mar 2023)","org":"OpenAI","src":"","value":125.9},{"date":"2024-01-25","model":"GPT-4 Turbo (Nov 2023)","org":"OpenAI","src":"","value":126.5},{"date":"2024-02-29","model":"Claude 3 Opus","org":"Anthropic","src":"","value":126.9},{"date":"2024-04-09","model":"GPT-4 Turbo (Apr 2024)","org":"OpenAI","src":"","value":127.2},{"date":"2024-05-13","model":"GPT-4o (May 2024)","org":"OpenAI","src":"","value":129},{"date":"2024-06-20","model":"Claude 3.5 Sonnet","org":"Anthropic","src":"","value":130},{"date":"2024-09-12","model":"o1-mini","org":"OpenAI","src":"","value":135.8},{"date":"2024-12-17","model":"o1","org":"OpenAI","src":"","value":141.9},{"date":"2025-03-31","model":"Gemini 2.5 Pro (Mar 2025)","org":"Google DeepMind","src":"","value":144.2},{"date":"2025-04-16","model":"o3","org":"OpenAI","src":"","value":146.9},{"date":"2025-06-10","model":"o3-pro","org":"OpenAI","src":"","value":147.4},{"date":"2025-08-07","model":"GPT-5","org":"OpenAI","src":"","value":150},{"date":"2025-10-07","model":"GPT-5 Pro","org":"OpenAI","src":"","value":150.3},{"date":"2025-11-18","model":"Gemini 3 Pro","org":"Google DeepMind","src":"","value":152.9},{"date":"2025-12-11","model":"GPT-5.2 Pro","org":"OpenAI","src":"","value":155.4},{"date":"2026-02-05","model":"GPT-5.3 Codex","org":"OpenAI","src":"","value":156.8},{"date":"2026-03-05","model":"GPT-5.4 Pro","org":"OpenAI","src":"","value":158.9},{"date":"2026-04-23","model":"GPT-5.5 Pro","org":"OpenAI","src":"","value":162.1},{"date":"2026-07-24","model":"Claude Opus 5","org":"Anthropic","src":"","value":162.8},{"date":"2026-09-01","model":"Claude Fable 5.1","org":"Anthropic","src":"","value":164.7},{"date":"2026-09-03","model":"GPT-6 Astra","org":"OpenAI","src":"","value":166.4},{"date":"2026-09-22","model":"Claude Opus 5.5","org":"Anthropic","src":"","value":167.3}],"group":"overall","introduced":"2025-10-01","key":"eci","last_model":"2026-09-29","measured":"epoch","models":274,"random":null,"results":274,"source":"Epoch AI","superseded":"","top":[{"date":"2026-09-22","model":"Claude Opus 5.5","org":"Anthropic","value":167.3},{"date":"2026-09-03","model":"GPT-6 Astra","org":"OpenAI","value":166.4},{"date":"2026-09-29","model":"GPT-6.1 Sol","org":"OpenAI","value":166.1},{"date":"2026-09-28","model":"Claude Sonnet 5.5","org":"Anthropic","value":165},{"date":"2026-09-01","model":"Claude Fable 5.1","org":"Anthropic","value":164.7},{"date":"2026-07-24","model":"Claude Opus 5","org":"Anthropic","value":162.8},{"date":"2026-09-22","model":"GPT-6 Sol","org":"OpenAI","value":162.7},{"date":"2026-04-23","model":"GPT-5.5 Pro","org":"OpenAI","value":162.1},{"date":"2026-06-09","model":"Claude Fable 5","org":"Anthropic","value":162.1},{"date":"2026-07-09","model":"GPT-5.6 Sol","org":"OpenAI","value":161.7}],"unit":"index"},{"active":true,"best":{"date":"2026-04-07","model":"Claude Mythos Preview (Early)","org":"Anthropic","src":"https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/","value":1044.78},"first_model":"2019-11-05","frontier":[{"date":"2019-11-05","model":"GPT-2 (1.5B)","org":"OpenAI","src":"https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/","value":0.05},{"date":"2023-03-14","model":"GPT-4 (Mar 2023)","org":"OpenAI","src":"METR - Measuring AI Ability to Complete Long Tasks","value":5.36},{"date":"2024-01-25","model":"GPT-4 Turbo Preview (January 2024)","org":"OpenAI","src":"METR - Measuring AI Ability to Complete Long Tasks","value":5.37},{"date":"2024-04-09","model":"GPT-4 Turbo (Apr 2024)","org":"OpenAI","src":"METR - Measuring AI Ability to Complete Long Tasks","value":6.57},{"date":"2024-05-13","model":"GPT-4o (May 2024)","org":"OpenAI","src":"METR - Measuring AI Ability to Complete Long Tasks","value":9.17},{"date":"2024-06-20","model":"Claude 3.5 Sonnet (Jun 2024)","org":"Anthropic","src":"https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/","value":11.4},{"date":"2024-09-12","model":"o1-preview","org":"OpenAI","src":"https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/","value":20.33},{"date":"2024-10-22","model":"Claude 3.5 Sonnet (Oct 2024)","org":"Anthropic","src":"https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/","value":20.52},{"date":"2024-12-17","model":"o1 (medium)","org":"OpenAI","src":"https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/","value":38.83},{"date":"2025-02-24","model":"Claude 3.7 Sonnet","org":"Anthropic","src":"","value":60.39},{"date":"2025-04-16","model":"o3 (unknown thinking)","org":"OpenAI","src":"","value":119.73},{"date":"2025-08-07","model":"GPT-5 (high)","org":"OpenAI","src":"","value":203.01},{"date":"2025-11-18","model":"Gemini 3 Pro Preview","org":"Google DeepMind","src":"","value":224.33},{"date":"2025-11-24","model":"Claude Opus 4.5 (no thinking)","org":"Anthropic","src":"","value":292.99},{"date":"2025-12-11","model":"GPT-5.2 (high)","org":"OpenAI","src":"","value":352.25},{"date":"2026-02-05","model":"Claude Opus 4.6 (unknown thinking)","org":"Anthropic","src":"","value":718.81},{"date":"2026-04-07","model":"Claude Mythos Preview (Early)","org":"Anthropic","src":"https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/","value":1044.78}],"group":"overall","introduced":"2025-03-19","key":"metr-time-horizon","last_model":"2026-04-07","measured":"external","models":39,"random":null,"results":49,"source":"METR","superseded":"","top":[{"date":"2026-04-07","model":"Claude Mythos Preview (Early)","org":"Anthropic","value":1044.78},{"date":"2026-02-05","model":"Claude Opus 4.6 (unknown thinking)","org":"Anthropic","value":718.81},{"date":"2026-02-19","model":"Gemini 3.1 Pro Preview","org":"Google DeepMind","value":384.15},{"date":"2025-12-11","model":"GPT-5.2 (high)","org":"OpenAI","value":352.25},{"date":"2026-02-05","model":"GPT-5.3 Codex","org":"OpenAI","value":349.53},{"date":"2026-03-05","model":"GPT-5.4 (xhigh)","org":"OpenAI","value":341.74},{"date":"2025-11-24","model":"Claude Opus 4.5 (no thinking)","org":"Anthropic","value":292.99},{"date":"2025-11-18","model":"Gemini 3 Pro Preview","org":"Google DeepMind","value":224.33},{"date":"2025-11-19","model":"GPT-5.1-Codex-Max","org":"OpenAI","value":223.71},{"date":"2025-08-07","model":"GPT-5 (high)","org":"OpenAI","value":203.01}],"unit":"minutes"},{"active":false,"best":{"date":"2024-11-20","model":"GPT-4o","org":"OpenAI","src":"Phi-4 Technical Report","value":88.1},"first_model":"2021-08-05","frontier":[{"date":"2021-08-05","model":"GPT-J-6B","org":"EleutherAI, LAION","src":"XGen-7B Technical Report","value":25.7},{"date":"2021-12-08","model":"Gopher","org":"DeepMind","src":"Scaling Language Models: Methods, Analysis \u0026 Insights from Training Gopher","value":60},{"date":"2022-03-15","model":"GPT-3.5","org":"OpenAI","src":"GPT-4 Technical Report","value":70},{"date":"2023-03-14","model":"GPT-4 (initial release)","org":"OpenAI","src":"GPT-4 Technical Report","value":86.4},{"date":"2024-06-20","model":"Claude 3.5 Sonnet (20240620)","org":"Anthropic","src":"Stanford CRFM Leaderboard","value":86.5},{"date":"2024-09-24","model":"Gemini 1.5 Pro (002)","org":"Google DeepMind","src":"Stanford CRFM Leaderboard","value":86.9},{"date":"2024-10-22","model":"Claude 3.5 Sonnet (20241022)","org":"Anthropic","src":"Stanford CRFM Leaderboard","value":87.3},{"date":"2024-11-20","model":"GPT-4o","org":"OpenAI","src":"Phi-4 Technical Report","value":88.1}],"group":"knowledge","introduced":"2020-09-07","key":"mmlu","last_model":"2024-12-26","measured":"external","models":158,"random":25,"results":249,"source":"papers and Stanford HELM, collected by Epoch AI","superseded":"","top":[{"date":"2024-11-20","model":"GPT-4o","org":"OpenAI","value":88.1},{"date":"2024-10-22","model":"Claude 3.5 Sonnet (20241022)","org":"Anthropic","value":87.3},{"date":"2024-12-26","model":"DeepSeek v3","org":"DeepSeek","value":87.2},{"date":"2024-09-24","model":"Gemini 1.5 Pro (002)","org":"Google DeepMind","value":86.9},{"date":"2024-06-20","model":"Claude 3.5 Sonnet (20240620)","org":"Anthropic","value":86.5},{"date":"2023-03-14","model":"GPT-4 (initial release)","org":"OpenAI","value":86.4},{"date":"2024-12-06","model":"Llama3.3 70b instruct","org":"Meta AI","value":86.3},{"date":"2024-05-14","model":"Gemini 1.5 Pro (May 2024)","org":"Google DeepMind","value":85.9},{"date":"2024-09-19","model":"Qwen2.5 72b instruct","org":"Alibaba","value":85.3},{"date":"2024-12-12","model":"Phi-4 14b","org":"Microsoft Research","value":84.8}],"unit":"percent"},{"active":true,"best":{"date":"2026-09-03","model":"GPT-6 Astra (max)","org":"OpenAI","src":"","value":95.8},"first_model":"2023-03-14","frontier":[{"date":"2023-03-14","model":"GPT-4 (Mar 2023)","org":"OpenAI","src":"","value":35.7},{"date":"2023-11-06","model":"GPT-4 Turbo Preview (Nov 2023)","org":"OpenAI","src":"","value":42.4},{"date":"2024-02-29","model":"Claude 3 Opus","org":"Anthropic","src":"","value":47.2},{"date":"2024-05-13","model":"GPT-4o (May 2024)","org":"OpenAI","src":"","value":48.9},{"date":"2024-06-20","model":"Claude 3.5 Sonnet (Jun 2024)","org":"Anthropic","src":"","value":54},{"date":"2024-09-12","model":"o1-mini (high)","org":"OpenAI","src":"","value":62.4},{"date":"2024-12-17","model":"o1 (high)","org":"OpenAI","src":"","value":76.8},{"date":"2025-01-31","model":"o3-mini (high)","org":"OpenAI","src":"","value":77},{"date":"2025-02-24","model":"Claude 3.7 Sonnet (64k thinking)","org":"Anthropic","src":"","value":79.7},{"date":"2025-03-25","model":"Gemini 2.5 Pro Exp (Mar 2025)","org":"Google DeepMind","src":"","value":83.8},{"date":"2025-06-05","model":"Gemini 2.5 Pro Preview (Jun 2025)","org":"Google DeepMind","src":"","value":84.8},{"date":"2025-07-09","model":"Grok 4","org":"xAI","src":"","value":87},{"date":"2025-11-13","model":"GPT-5.1 (high)","org":"OpenAI","src":"","value":87.6},{"date":"2025-11-18","model":"Gemini 3 Pro Preview","org":"Google DeepMind","src":"","value":92.6},{"date":"2026-02-19","model":"Gemini 3.1 Pro Preview","org":"Google DeepMind","src":"","value":94.4},{"date":"2026-03-05","model":"GPT-5.4 Pro (xhigh)","org":"OpenAI","src":"","value":94.6},{"date":"2026-08-13","model":"Gemini 3.7 Flash (high)","org":"Google DeepMind","src":"","value":94.8},{"date":"2026-09-02","model":"Gemini 3.8 Flash (high)","org":"Google DeepMind","src":"","value":95.4},{"date":"2026-09-03","model":"GPT-6 Astra (max)","org":"OpenAI","src":"","value":95.8}],"group":"knowledge","introduced":"2023-11-20","key":"gpqa-diamond","last_model":"2026-10-07","measured":"epoch","models":214,"random":25,"results":321,"source":"Epoch AI (own runs)","superseded":"","top":[{"date":"2026-09-03","model":"GPT-6 Astra (max)","org":"OpenAI","value":95.8},{"date":"2026-09-28","model":"Claude Sonnet 5.5 (max)","org":"Anthropic","value":95.6},{"date":"2026-09-02","model":"Gemini 3.8 Flash (high)","org":"Google DeepMind","value":95.4},{"date":"2026-09-29","model":"GPT-6.1 Sol (max)","org":"OpenAI","value":95.4},{"date":"2026-08-13","model":"Gemini 3.7 Flash (high)","org":"Google DeepMind","value":94.8},{"date":"2026-03-05","model":"GPT-5.4 Pro (xhigh)","org":"OpenAI","value":94.6},{"date":"2026-02-19","model":"Gemini 3.1 Pro Preview","org":"Google DeepMind","value":94.4},{"date":"2026-09-22","model":"GPT-6 Sol (max)","org":"OpenAI","value":94.3},{"date":"2026-07-21","model":"Gemini 3.6 Flash","org":"Google DeepMind","value":94.1},{"date":"2026-04-23","model":"GPT-5.5 (xhigh)","org":"OpenAI","value":94}],"unit":"percent"},{"active":true,"best":{"date":"2026-09-03","model":"GPT 6 Astra","org":"OpenAI","src":"","value":54.8},"first_model":"2024-09-24","frontier":[{"date":"2024-09-24","model":"Gemini-1.5-Pro-002","org":"Google DeepMind","src":"","value":4.6},{"date":"2024-12-17","model":"o1 (December 2024)","org":"OpenAI","src":"","value":8},{"date":"2025-03-19","model":"o1 Pro","org":"OpenAI","src":"","value":8.1},{"date":"2025-03-25","model":"Gemini 2.5 Pro Experimental (March 2025)","org":"Google DeepMind","src":"","value":18.2},{"date":"2025-04-16","model":"o3 (high) (April 2025)","org":"OpenAI","src":"","value":20.3},{"date":"2025-06-05","model":"gemini-2.5-pro-preview-06-05","org":"Google DeepMind","src":"","value":21.6},{"date":"2025-08-07","model":"gpt-5-2025-08-07","org":"OpenAI","src":"","value":25.3},{"date":"2025-10-07","model":"gpt-5-pro-2025-10-06","org":"OpenAI","src":"","value":31.6},{"date":"2025-11-18","model":"gemini-3-pro-preview","org":"Google DeepMind","src":"","value":37.5},{"date":"2026-02-19","model":"Gemini 3.1 Pro Preview (thinking high)","org":"Google DeepMind","src":"","value":46.4},{"date":"2026-09-01","model":"Fable 5.1 (xhigh)","org":"Anthropic","src":"","value":46.5},{"date":"2026-09-03","model":"GPT 6 Astra","org":"OpenAI","src":"","value":54.8}],"group":"knowledge","introduced":"2025-01-23","key":"hle","last_model":"2026-09-03","measured":"external","models":44,"random":4.8,"results":54,"source":"Humanity's Last Exam leaderboard, collected by Epoch AI","superseded":"","top":[{"date":"2026-09-03","model":"GPT 6 Astra","org":"OpenAI","value":54.8},{"date":"2026-09-01","model":"Fable 5.1 (xhigh)","org":"Anthropic","value":46.5},{"date":"2026-02-19","model":"Gemini 3.1 Pro Preview (thinking high)","org":"Google DeepMind","value":46.4},{"date":"2026-09-02","model":"Gemini 3.8 Flash","org":"Google DeepMind","value":44.5},{"date":"2026-03-05","model":"GPT-5.4 Pro","org":"OpenAI","value":44.3},{"date":"2026-04-08","model":"Muse Spark","org":"Meta AI","value":40.6},{"date":"2025-11-18","model":"gemini-3-pro-preview","org":"Google DeepMind","value":37.5},{"date":"2026-03-05","model":"gpt-5.4-2026-03-05 (xhigh thinking)","org":"OpenAI","value":36.2},{"date":"2026-04-16","model":"Claude Opus 4.7","org":"Anthropic","value":36.2},{"date":"2026-02-05","model":"claude-opus-4-6-thinking-max","org":"Anthropic","value":34.4}],"unit":"percent"},{"active":true,"best":{"date":"2026-09-03","model":"GPT-6 Astra (max)","org":"OpenAI","src":"","value":75.6},"first_model":"2024-02-29","frontier":[{"date":"2024-02-29","model":"Claude 3 Opus","org":"Anthropic","src":"","value":12.6},{"date":"2024-08-06","model":"GPT-4o (Aug 2024)","org":"OpenAI","src":"","value":26},{"date":"2024-12-17","model":"o1 (high)","org":"OpenAI","src":"","value":41.1},{"date":"2025-04-16","model":"o3 (high)","org":"OpenAI","src":"","value":49.4},{"date":"2025-08-07","model":"GPT-5 (high)","org":"OpenAI","src":"","value":50.1},{"date":"2025-12-17","model":"Gemini 3 Flash Preview","org":"Google DeepMind","src":"","value":66.8},{"date":"2026-02-19","model":"Gemini 3.1 Pro Preview","org":"Google DeepMind","src":"","value":73.5},{"date":"2026-09-03","model":"GPT-6 Astra (max)","org":"OpenAI","src":"","value":75.6}],"group":"knowledge","introduced":"2025-09-09","key":"simpleqa-verified","last_model":"2026-10-07","measured":"epoch","models":79,"random":null,"results":88,"source":"Epoch AI (own runs)","superseded":"","top":[{"date":"2026-09-03","model":"GPT-6 Astra (max)","org":"OpenAI","value":75.6},{"date":"2026-09-29","model":"GPT-6.1 Sol (max)","org":"OpenAI","value":73.9},{"date":"2026-02-19","model":"Gemini 3.1 Pro Preview","org":"Google DeepMind","value":73.5},{"date":"2026-09-22","model":"Claude Opus 5.5 (max)","org":"Anthropic","value":72.2},{"date":"2026-09-01","model":"Claude Fable 5.1 (max)","org":"Anthropic","value":70.8},{"date":"2026-06-09","model":"Claude Fable 5 (xhigh)","org":"Anthropic","value":70.7},{"date":"2026-07-09","model":"GPT-5.6 Sol (max)","org":"OpenAI","value":69.7},{"date":"2026-09-02","model":"Gemini 3.8 Flash (high)","org":"Google DeepMind","value":69.7},{"date":"2026-08-13","model":"Gemini 3.7 Flash (high)","org":"Google DeepMind","value":69.2},{"date":"2025-12-17","model":"Gemini 3 Flash Preview","org":"Google DeepMind","value":66.8}],"unit":"percent"},{"active":false,"best":{"date":"2024-06-17","model":"DS-Coder-V2-Instruct","org":"DeepSeek","src":"Qwen2.5-Coder Technical Report","value":94.5},"first_model":"2020-06-22","frontier":[{"date":"2020-06-22","model":"babbage (1.3B)","org":"OpenAI","src":"Stanford HELM","value":0.7},{"date":"2022-03-15","model":"code-davinci-002","org":"OpenAI","src":"Stanford HELM","value":56.8},{"date":"2022-11-28","model":"GPT-3.5 (8-shot)","org":"OpenAI","src":"Qwen Technical Report","value":78.2},{"date":"2023-03-14","model":"GPT-4","org":"OpenAI","src":"GPT-4 Technical Report","value":92},{"date":"2024-06-17","model":"DS-Coder-V2-Instruct","org":"DeepSeek","src":"Qwen2.5-Coder Technical Report","value":94.5}],"group":"math","introduced":"2021-10-27","key":"gsm8k","last_model":"2024-11-29","measured":"external","models":146,"random":null,"results":235,"source":"papers and Stanford HELM, collected by Epoch AI","superseded":"","top":[{"date":"2024-06-17","model":"DS-Coder-V2-Instruct","org":"DeepSeek","value":94.5},{"date":"","model":"Qwen2.5-Coder-14B-Instruct","org":"","value":94.2},{"date":"2024-11-21","model":"Qwen2.5-Coder-32B-Instruct","org":"Alibaba","value":93},{"date":"2023-03-14","model":"GPT-4","org":"OpenAI","value":92},{"date":"","model":"GPT-4","org":"","value":92},{"date":"","model":"GPT-4 (8-shot)","org":"","value":91.4},{"date":"2024-07-18","model":"GPT-4o-mini","org":"OpenAI","value":91.3},{"date":"2024-09-18","model":"Qwen2.5-Coder-32B","org":"Alibaba","value":91.1},{"date":"2023-06-13","model":"GPT-4","org":"OpenAI","value":90},{"date":"2024-08-17","model":"Phi-3.5-MoE","org":"Microsoft","value":88.7}],"unit":"percent"},{"active":true,"best":{"date":"2025-08-07","model":"GPT-5 (high)","org":"OpenAI","src":"","value":98.1},"first_model":"2023-06-13","frontier":[{"date":"2023-06-13","model":"GPT-4 (Jun 2023)","org":"OpenAI","src":"","value":23},{"date":"2023-11-06","model":"GPT-4 Turbo Preview (Nov 2023)","org":"OpenAI","src":"","value":40},{"date":"2024-04-09","model":"GPT-4 Turbo (Apr 2024)","org":"OpenAI","src":"","value":46.7},{"date":"2024-05-13","model":"GPT-4o (May 2024)","org":"OpenAI","src":"","value":51},{"date":"2024-06-20","model":"Claude 3.5 Sonnet (Jun 2024)","org":"Anthropic","src":"","value":51.7},{"date":"2024-07-18","model":"GPT-4o mini","org":"OpenAI","src":"","value":52.6},{"date":"2024-08-06","model":"GPT-4o (Aug 2024)","org":"OpenAI","src":"","value":53.3},{"date":"2024-09-12","model":"o1-mini (high)","org":"OpenAI","src":"","value":89.2},{"date":"2024-12-17","model":"o1 (high)","org":"OpenAI","src":"","value":94.7},{"date":"2025-01-31","model":"o3-mini (high)","org":"OpenAI","src":"","value":96.5},{"date":"2025-04-16","model":"o4-mini (high)","org":"OpenAI","src":"","value":97.8},{"date":"2025-08-07","model":"GPT-5 (high)","org":"OpenAI","src":"","value":98.1}],"group":"math","introduced":"2021-03-05","key":"math-level-5","last_model":"2025-10-15","measured":"epoch","models":94,"random":null,"results":108,"source":"Epoch AI (own runs)","superseded":"","top":[{"date":"2025-08-07","model":"GPT-5 (high)","org":"OpenAI","value":98.1},{"date":"2025-04-16","model":"o4-mini (high)","org":"OpenAI","value":97.8},{"date":"2025-04-16","model":"o3 (high)","org":"OpenAI","value":97.8},{"date":"2025-08-07","model":"GPT-5 mini (high)","org":"OpenAI","value":97.8},{"date":"2025-09-29","model":"Claude Sonnet 4.5 (32k thinking)","org":"Anthropic","value":97.7},{"date":"2025-09-24","model":"Qwen3-Max-Instruct","org":"Alibaba","value":97.1},{"date":"2025-05-28","model":"DeepSeek-R1 (May 2025)","org":"DeepSeek","value":96.6},{"date":"2025-01-31","model":"o3-mini (high)","org":"OpenAI","value":96.5},{"date":"2025-10-15","model":"Claude Haiku 4.5","org":"Anthropic","value":96.4},{"date":"2025-05-06","model":"Gemini 2.5 Pro Preview (Jun 2025)","org":"Google DeepMind","value":95.9}],"unit":"percent"},{"active":true,"best":{"date":"2026-09-03","model":"GPT-6 Astra (max)","org":"OpenAI","src":"","value":93.7},"first_model":"2024-01-25","frontier":[{"date":"2024-01-25","model":"GPT-3.5 Turbo (Jan 2024)","org":"OpenAI","src":"","value":0},{"date":"2024-04-09","model":"GPT-4 Turbo (Apr 2024)","org":"OpenAI","src":"","value":0.7},{"date":"2024-12-17","model":"o1 (high)","org":"OpenAI","src":"","value":14.7},{"date":"2025-01-31","model":"o3-mini (high)","org":"OpenAI","src":"","value":18.6},{"date":"2025-04-16","model":"o4-mini (high)","org":"OpenAI","src":"","value":36.1},{"date":"2025-08-07","model":"GPT-5 (high)","org":"OpenAI","src":"","value":55.4},{"date":"2025-10-07","model":"GPT-5 Pro","org":"OpenAI","src":"","value":55.8},{"date":"2025-12-11","model":"GPT-5.2 Pro","org":"OpenAI","src":"","value":74},{"date":"2026-03-05","model":"GPT-5.4 Pro (xhigh)","org":"OpenAI","src":"","value":82.5},{"date":"2026-04-23","model":"GPT-5.5 Pro (xhigh)","org":"OpenAI","src":"","value":87.7},{"date":"2026-07-09","model":"GPT-5.6 Sol (max)","org":"OpenAI","src":"","value":89.1},{"date":"2026-09-01","model":"Claude Fable 5.1 (max)","org":"Anthropic","src":"","value":90.2},{"date":"2026-09-03","model":"GPT-6 Astra (max)","org":"OpenAI","src":"","value":93.7}],"group":"math","introduced":"2026-06-12","key":"frontiermath","last_model":"2026-10-07","measured":"epoch","models":83,"random":null,"results":116,"source":"Epoch AI (own runs)","superseded":"","top":[{"date":"2026-09-03","model":"GPT-6 Astra (max)","org":"OpenAI","value":93.7},{"date":"2026-09-29","model":"GPT-6.1 Sol (max)","org":"OpenAI","value":93.7},{"date":"2026-09-22","model":"Claude Opus 5.5 (max)","org":"Anthropic","value":91.2},{"date":"2026-09-01","model":"Claude Fable 5.1 (max)","org":"Anthropic","value":90.2},{"date":"2026-09-22","model":"GPT-6 Sol (max)","org":"OpenAI","value":89.8},{"date":"2026-07-09","model":"GPT-5.6 Sol (max)","org":"OpenAI","value":89.1},{"date":"2026-09-28","model":"Claude Sonnet 5.5 (max)","org":"Anthropic","value":88.8},{"date":"2026-04-23","model":"GPT-5.5 Pro (xhigh)","org":"OpenAI","value":87.7},{"date":"2026-06-09","model":"Claude Fable 5 (max)","org":"Anthropic","value":87},{"date":"2026-07-09","model":"GPT-5.6 Terra (max)","org":"OpenAI","value":86}],"unit":"percent"},{"active":true,"best":{"date":"2026-04-16","model":"Claude Opus 4.7 (max)","org":"Anthropic","src":"","value":83.5},"first_model":"2024-11-20","frontier":[{"date":"2024-11-20","model":"GPT-4o (Nov 2024)","org":"OpenAI","src":"","value":31},{"date":"2025-02-24","model":"Claude 3.7 Sonnet","org":"Anthropic","src":"","value":61},{"date":"2025-04-16","model":"o3 (medium)","org":"OpenAI","src":"","value":62.3},{"date":"2025-05-22","model":"Claude Opus 4","org":"Anthropic","src":"","value":70.7},{"date":"2025-08-05","model":"Claude Opus 4.1","org":"Anthropic","src":"","value":73.3},{"date":"2025-08-07","model":"GPT-5 (high)","org":"OpenAI","src":"","value":73.6},{"date":"2025-11-24","model":"Claude Opus 4.5 (no thinking)","org":"Anthropic","src":"","value":76.7},{"date":"2026-02-05","model":"Claude Opus 4.6 (no thinking)","org":"Anthropic","src":"","value":78.7},{"date":"2026-04-16","model":"Claude Opus 4.7 (max)","org":"Anthropic","src":"","value":83.5}],"group":"coding","introduced":"2024-08-13","key":"swe-bench-verified","last_model":"2026-06-16","measured":"epoch","models":32,"random":null,"results":35,"source":"Epoch AI (own runs)","superseded":"","top":[{"date":"2026-04-16","model":"Claude Opus 4.7 (max)","org":"Anthropic","value":83.5},{"date":"2026-04-23","model":"GPT-5.5 (xhigh)","org":"OpenAI","value":80.6},{"date":"2026-05-19","model":"Gemini 3.5 Flash","org":"Google DeepMind","value":79.3},{"date":"2026-02-05","model":"Claude Opus 4.6 (no thinking)","org":"Anthropic","value":78.7},{"date":"2026-06-16","model":"GLM-5.2","org":"Z.ai (Zhipu AI)","value":78.7},{"date":"2026-04-24","model":"DeepSeek v4 Pro (max)","org":"DeepSeek","value":77.6},{"date":"2026-05-19","model":"Qwen3.7 Max","org":"Alibaba","value":77.3},{"date":"2026-03-05","model":"GPT-5.4 (high)","org":"OpenAI","value":76.9},{"date":"2025-11-24","model":"Claude Opus 4.5 (no thinking)","org":"Anthropic","value":76.7},{"date":"2026-04-20","model":"Qwen 3.6 Max (Preview)","org":"Alibaba","value":76.7}],"unit":"percent"},{"active":true,"best":{"date":"2025-08-07","model":"GPT-5 (high)","org":"OpenAI","src":"https://aider.chat/docs/leaderboards/","value":88},"first_model":"2024-07-18","frontier":[{"date":"2024-07-18","model":"GPT-4o mini","org":"OpenAI","src":"https://aider.chat/docs/leaderboards/","value":3.6},{"date":"2024-08-06","model":"GPT-4o (Aug 2024)","org":"OpenAI","src":"https://aider.chat/docs/leaderboards/","value":23.1},{"date":"2024-09-12","model":"o1-mini (unknown thinking)","org":"OpenAI","src":"https://aider.chat/docs/leaderboards/","value":32.9},{"date":"2024-10-22","model":"Claude 3.5 Sonnet (Oct 2024)","org":"Anthropic","src":"https://aider.chat/docs/leaderboards/","value":51.6},{"date":"2024-12-17","model":"o1 (high)","org":"OpenAI","src":"https://aider.chat/docs/leaderboards/","value":61.7},{"date":"2025-02-24","model":"Claude 3.7 Sonnet (32k thinking)","org":"Anthropic","src":"https://aider.chat/docs/leaderboards/","value":64.9},{"date":"2025-03-25","model":"Gemini 2.5 Pro Exp (Mar 2025)","org":"Google DeepMind","src":"Aider LLM Leaderboards","value":72.9},{"date":"2025-04-16","model":"o3 (high)","org":"OpenAI","src":"https://aider.chat/docs/leaderboards/","value":81.3},{"date":"2025-06-05","model":"Gemini 2.5 Pro Preview (Jun 2025)","org":"Google DeepMind","src":"https://aider.chat/docs/leaderboards/","value":83.1},{"date":"2025-06-10","model":"o3-pro","org":"OpenAI","src":"https://aider.chat/docs/leaderboards/","value":84.9},{"date":"2025-08-07","model":"GPT-5 (high)","org":"OpenAI","src":"https://aider.chat/docs/leaderboards/","value":88}],"group":"coding","introduced":"2024-12-21","key":"aider-polyglot","last_model":"2025-12-01","measured":"external","models":54,"random":null,"results":72,"source":"Aider leaderboard, collected by Epoch AI","superseded":"","top":[{"date":"2025-08-07","model":"GPT-5 (high)","org":"OpenAI","value":88},{"date":"2025-06-10","model":"o3-pro","org":"OpenAI","value":84.9},{"date":"2025-06-05","model":"Gemini 2.5 Pro Preview (Jun 2025)","org":"Google DeepMind","value":83.1},{"date":"2025-04-16","model":"o3 (high)","org":"OpenAI","value":81.3},{"date":"2025-07-09","model":"Grok 4","org":"xAI","value":79.6},{"date":"2025-05-06","model":"Gemini 2.5 Pro Preview (Jun 2025)","org":"Google DeepMind","value":76.9},{"date":"2025-09-29","model":"DeepSeek-V3.2-Exp","org":"DeepSeek","value":74.2},{"date":"2025-12-01","model":"DeepSeek-V3.2 (Thinking)","org":"DeepSeek","value":74.2},{"date":"2025-03-25","model":"Gemini 2.5 Pro Exp (Mar 2025)","org":"Google DeepMind","value":72.9},{"date":"2025-04-16","model":"o4-mini (high)","org":"OpenAI","value":72}],"unit":"percent"},{"active":true,"best":{"date":"2026-04-23","model":"GPT-5.5","org":"OpenAI","src":"https://www.tbench.ai/leaderboard/terminal-bench/2.0","value":84.7},"first_model":"2025-06-17","frontier":[{"date":"2025-06-17","model":"Gemini 2.5 Flash","org":"Google DeepMind","src":"https://www.tbench.ai/leaderboard/terminal-bench/2.0","value":17.1},{"date":"2025-07-09","model":"Grok 4","org":"xAI","src":"Terminal-Bench v2 Leaderboard","value":27.2},{"date":"2025-07-12","model":"Kimi K2 Instruct","org":"Moonshot","src":"Terminal-Bench v2 Leaderboard","value":27.8},{"date":"2025-08-05","model":"Claude Opus 4.1","org":"Anthropic","src":"Terminal-Bench v2 Leaderboard","value":38},{"date":"2025-08-07","model":"GPT-5","org":"OpenAI","src":"https://www.tbench.ai/leaderboard/terminal-bench/2.0","value":49.6},{"date":"2025-11-12","model":"GPT-5.1-Codex-Mini","org":"OpenAI","src":"https://www.tbench.ai/leaderboard/terminal-bench/2.0","value":61.6},{"date":"2025-11-18","model":"Gemini 3 Pro","org":"Google DeepMind","src":"https://www.tbench.ai/leaderboard/terminal-bench/2.0","value":69.4},{"date":"2026-02-05","model":"Claude Opus 4.6","org":"Anthropic","src":"https://www.tbench.ai/leaderboard/terminal-bench/2.0","value":79.8},{"date":"2026-02-19","model":"Gemini 3.1 Pro","org":"Google DeepMind","src":"https://www.tbench.ai/leaderboard/terminal-bench/2.0","value":80.2},{"date":"2026-03-05","model":"GPT-5.4","org":"OpenAI","src":"https://www.tbench.ai/leaderboard/terminal-bench/2.0","value":81.8},{"date":"2026-04-23","model":"GPT-5.5","org":"OpenAI","src":"https://www.tbench.ai/leaderboard/terminal-bench/2.0","value":84.7}],"group":"coding","introduced":"2025-05-19","key":"terminal-bench","last_model":"2026-04-23","measured":"external","models":45,"random":null,"results":202,"source":"Terminal-Bench leaderboard, collected by Epoch AI","superseded":"","top":[{"date":"2026-04-23","model":"GPT-5.5","org":"OpenAI","value":84.7},{"date":"2026-03-05","model":"GPT-5.4","org":"OpenAI","value":81.8},{"date":"2026-02-19","model":"Gemini 3.1 Pro","org":"Google DeepMind","value":80.2},{"date":"2026-04-16","model":"Claude Opus 4.7","org":"Anthropic","value":80.2},{"date":"2026-02-05","model":"Claude Opus 4.6","org":"Anthropic","value":79.8},{"date":"2026-02-05","model":"GPT-5.3-Codex","org":"OpenAI","value":78.4},{"date":"2025-11-18","model":"Gemini 3 Pro","org":"Google DeepMind","value":69.4},{"date":"2025-12-18","model":"GPT-5.2-Codex","org":"OpenAI","value":66.5},{"date":"2025-12-11","model":"GPT-5.2","org":"OpenAI","value":64.9},{"date":"2025-12-17","model":"Gemini 3 Flash","org":"Google DeepMind","value":64.3}],"unit":"percent"},{"active":true,"best":{"date":"2026-06-09","model":"Claude Fable 5 (XHigh)","org":"Anthropic","src":"https://arcprize.org/leaderboard","value":98.5},"first_model":"2024-09-12","frontier":[{"date":"2024-09-12","model":"o1-preview","org":"OpenAI","src":"ARC Prize Leaderboard","value":18},{"date":"2024-12-17","model":"o1 (Medium)","org":"OpenAI","src":"ARC Prize Leaderboard","value":30.7},{"date":"2025-01-31","model":"o3-mini (High)","org":"OpenAI","src":"ARC Prize Leaderboard","value":34.5},{"date":"2025-04-16","model":"o3 (High)","org":"OpenAI","src":"https://arcprize.org/leaderboard","value":60.8},{"date":"2025-07-09","model":"Grok 4 (Thinking)","org":"xAI","src":"","value":66.7},{"date":"2025-10-07","model":"GPT-5 Pro","org":"OpenAI","src":"ARC Prize Leaderboard","value":70.2},{"date":"2025-11-13","model":"GPT-5.1 (Thinking, High)","org":"OpenAI","src":"https://arcprize.org/leaderboard","value":72.8},{"date":"2025-11-18","model":"Gemini 3 Pro","org":"Google DeepMind","src":"ARC Prize Leaderboard","value":75},{"date":"2025-11-24","model":"Claude Opus 4.5 (64k thinking)","org":"Anthropic","src":"ARC Prize Leaderboard","value":80},{"date":"2025-12-11","model":"GPT-5.2 Pro (X-High)","org":"OpenAI","src":"","value":90.5},{"date":"2026-02-05","model":"Claude Opus 4.6 (120K, High)","org":"Anthropic","src":"https://arcprize.org/leaderboard","value":94},{"date":"2026-02-12","model":"Gemini 3 Deep Think (2/26)","org":"Google DeepMind","src":"https://arcprize.org/leaderboard","value":96},{"date":"2026-02-19","model":"Gemini 3.1 Pro (Preview)","org":"Google DeepMind","src":"ARC Prize Leaderboard","value":98},{"date":"2026-06-09","model":"Claude Fable 5 (XHigh)","org":"Anthropic","src":"https://arcprize.org/leaderboard","value":98.5}],"group":"abstraction","introduced":"2019-11-05","key":"arc-agi","last_model":"2026-09-29","measured":"external","models":91,"random":null,"results":279,"source":"ARC Prize leaderboard, collected by Epoch AI","superseded":"","top":[{"date":"2026-06-09","model":"Claude Fable 5 (XHigh)","org":"Anthropic","value":98.5},{"date":"2026-09-02","model":"Gemini 3.8 Flash (High)","org":"Google DeepMind","value":98.5},{"date":"2026-09-03","model":"GPT-6 Astra (XHigh)","org":"OpenAI","value":98.5},{"date":"2026-09-22","model":"Claude Opus 5.5 (High)","org":"Anthropic","value":98.5},{"date":"2026-09-29","model":"GPT-6.1 Sol (XHigh)","org":"OpenAI","value":98.5},{"date":"2026-02-19","model":"Gemini 3.1 Pro (Preview)","org":"Google DeepMind","value":98},{"date":"2026-07-09","model":"GPT-5.6 Sol (xHigh)","org":"OpenAI","value":97.5},{"date":"2026-07-24","model":"Claude Opus 5 (High)","org":"Anthropic","value":97.5},{"date":"2026-09-01","model":"Claude Fable 5.1 (Max)","org":"Anthropic","value":97.5},{"date":"2026-04-23","model":"GPT-5.5 Pro (High)","org":"OpenAI","value":96.5}],"unit":"percent"},{"active":true,"best":{"date":"2026-09-03","model":"GPT-6 Astra (Max)","org":"OpenAI","src":"","value":95},"first_model":"2024-07-18","frontier":[{"date":"2024-07-18","model":"GPT-4o-mini","org":"OpenAI","src":"","value":0},{"date":"2024-09-12","model":"o1-mini","org":"OpenAI","src":"","value":0.8},{"date":"2025-01-20","model":"Deepseek R1","org":"DeepSeek","src":"","value":1.3},{"date":"2025-01-31","model":"o3-mini (High)","org":"OpenAI","src":"","value":3},{"date":"2025-04-16","model":"o3 (High)","org":"OpenAI","src":"","value":6.5},{"date":"2025-05-22","model":"Claude Opus 4 (Thinking 16K)","org":"Anthropic","src":"","value":8.6},{"date":"2025-07-09","model":"Grok 4 (Thinking)","org":"xAI","src":"","value":16},{"date":"2025-10-07","model":"GPT-5 Pro","org":"OpenAI","src":"","value":18.3},{"date":"2025-11-18","model":"Gemini 3 Pro","org":"Google DeepMind","src":"","value":31.1},{"date":"2025-11-24","model":"Opus 4.5 (Thinking, 64K)","org":"Anthropic","src":"","value":37.6},{"date":"2025-12-11","model":"GPT-5.2 Pro (High)","org":"OpenAI","src":"","value":54.2},{"date":"2026-02-05","model":"Claude Opus 4.6 (120K, High)","org":"Anthropic","src":"","value":69.2},{"date":"2026-02-12","model":"Gemini 3 Deep Think (2/26)","org":"Google DeepMind","src":"","value":84.6},{"date":"2026-04-23","model":"GPT-5.5 (xHigh)","org":"OpenAI","src":"","value":85},{"date":"2026-06-09","model":"Claude Fable 5 (Max)","org":"Anthropic","src":"","value":89.2},{"date":"2026-07-09","model":"GPT-5.6 Sol (Max)","org":"OpenAI","src":"","value":92.5},{"date":"2026-09-03","model":"GPT-6 Astra (Max)","org":"OpenAI","src":"","value":95}],"group":"abstraction","introduced":"2025-03-24","key":"arc-agi-2","last_model":"2026-09-29","measured":"external","models":86,"random":null,"results":259,"source":"ARC Prize leaderboard, collected by Epoch AI","superseded":"","top":[{"date":"2026-09-03","model":"GPT-6 Astra (Max)","org":"OpenAI","value":95},{"date":"2026-09-29","model":"GPT-6.1 Sol (Max)","org":"OpenAI","value":94.2},{"date":"2026-09-22","model":"Claude Opus 5.5 (High)","org":"Anthropic","value":93.3},{"date":"2026-07-09","model":"GPT-5.6 Sol (Max)","org":"OpenAI","value":92.5},{"date":"2026-07-24","model":"Claude Opus 5 (Max)","org":"Anthropic","value":90.4},{"date":"2026-09-01","model":"Claude Fable 5.1 (Max)","org":"Anthropic","value":90},{"date":"2026-09-22","model":"GPT-6 Sol (Max)","org":"OpenAI","value":89.6},{"date":"2026-06-09","model":"Claude Fable 5 (Max)","org":"Anthropic","value":89.2},{"date":"2026-09-02","model":"Gemini 3.8 Flash (High)","org":"Google DeepMind","value":89.2},{"date":"2026-04-23","model":"GPT-5.5 (xHigh)","org":"OpenAI","value":85}],"unit":"percent"},{"active":false,"best":{"date":"2023-03-14","model":"GPT-4","org":"OpenAI","src":"The Falcon Series of Open Language Models","value":95.3},"first_model":"2019-11-05","frontier":[{"date":"2019-11-05","model":"GPT2-XL-1.5B","org":"OpenAI","src":"Textbooks Are All You Need II: phi-1.5 technical report","value":40},{"date":"2020-06-22","model":"babbage (1.3B)","org":"OpenAI","src":"Stanford HELM","value":55.5},{"date":"2021-08-05","model":"GPT-J-6B","org":"EleutherAI, LAION","src":"XGen-7B Technical Report","value":66.2},{"date":"2021-10-11","model":"Megatron-NLG (530B)","org":"Microsoft, Nvidia","src":"GLaM: Efficient Scaling of Language Models with Mixture-of-Experts","value":82.4},{"date":"2022-04-04","model":"PaLM-540B","org":"Google Research","src":"PaLM: Scaling Language Modeling with Pathways","value":83.8},{"date":"2022-11-28","model":"GPT-3.5","org":"OpenAI","src":"The Falcon Series of Open Language Models","value":85.5},{"date":"2023-03-14","model":"GPT-4","org":"OpenAI","src":"The Falcon Series of Open Language Models","value":95.3}],"group":"language","introduced":"2019-05-19","key":"hellaswag","last_model":"2024-12-26","measured":"external","models":97,"random":25,"results":135,"source":"papers and Stanford HELM, collected by Epoch AI","superseded":"","top":[{"date":"2023-03-14","model":"GPT-4","org":"OpenAI","value":95.3},{"date":"2024-07-23","model":"LLaMA-3.1 405B Base","org":"Meta AI","value":89.2},{"date":"2023-09-06","model":"Falcon-180B","org":"Technology Innovation Institute","value":89},{"date":"2024-12-26","model":"DeepSeek-V3 Base","org":"DeepSeek","value":88.9},{"date":"2024-05-07","model":"DeepSeek-V2 Base","org":"DeepSeek","value":87.1},{"date":"","model":"PaLM 2-L","org":"","value":86.8},{"date":"2023-12-11","model":"Mixtral 8x7B","org":"Mistral AI","value":86.7},{"date":"","model":"UNICORN","org":"","value":86.6},{"date":"2022-11-28","model":"GPT-3.5","org":"OpenAI","value":85.5},{"date":"2023-03-15","model":"Falcon 40B","org":"Technology Innovation Institute","value":85.3}],"unit":"percent"},{"active":false,"best":{"date":"2024-07-23","model":"LLaMA-3.1 405B Base","org":"Meta AI","src":"DeepSeek-V3 Technical Report","value":89.2},"first_model":"2019-11-05","frontier":[{"date":"2019-11-05","model":"GPT-2-XL 1.5B","org":"OpenAI","src":"","value":58.3},{"date":"2021-08-05","model":"GPT-J-6B","org":"EleutherAI, LAION","src":"XGen-7B Technical Report","value":64.5},{"date":"2021-10-11","model":"Megatron-NLG (530B)","org":"Microsoft, Nvidia","src":"GLaM: Efficient Scaling of Language Models with Mixture-of-Experts","value":78.9},{"date":"2021-12-13","model":"GLaM","org":"Google","src":"GLaM: Efficient Scaling of Language Models with Mixture-of-Experts","value":79.2},{"date":"2022-03-15","model":"GPT-3.5","org":"OpenAI","src":"GPT-4 technical report","value":81.6},{"date":"2022-04-04","model":"5-shot PaLM","org":"Google Research","src":"GPT-4 Technical report","value":85.1},{"date":"2023-03-14","model":"GPT-4","org":"OpenAI","src":"GPT-4 technical report","value":87.5},{"date":"2024-02-29","model":"Claude 3 Opus","org":"Anthropic","src":"The Claude 3 Model Family: Opus, Sonnet, Haiku","value":88.5},{"date":"2024-07-23","model":"LLaMA-3.1 405B Base","org":"Meta AI","src":"DeepSeek-V3 Technical Report","value":89.2}],"group":"language","introduced":"2019-07-24","key":"winogrande","last_model":"2024-12-26","measured":"external","models":84,"random":50,"results":145,"source":"papers, collected by Epoch AI","superseded":"","top":[{"date":"2024-07-23","model":"LLaMA-3.1 405B Base","org":"Meta AI","value":89.2},{"date":"2024-02-29","model":"Claude 3 Opus","org":"Anthropic","value":88.5},{"date":"2023-03-14","model":"GPT-4","org":"OpenAI","value":87.5},{"date":"2023-09-06","model":"Falcon-180B","org":"Technology Innovation Institute","value":87.1},{"date":"","model":"UNICORN","org":"","value":86.6},{"date":"2024-05-07","model":"DeepSeek-V2 Base","org":"DeepSeek","value":86.3},{"date":"2024-12-26","model":"DeepSeek-V3 Base","org":"DeepSeek","value":85.2},{"date":"2022-04-04","model":"5-shot PaLM","org":"Google Research","value":85.1},{"date":"2024-06-17","model":"DS-Coder-V2-Base","org":"DeepSeek","value":83.7},{"date":"2024-04-18","model":"LLaMA 3 70B","org":"Meta AI","value":83.5}],"unit":"percent"}],"license":"CC BY 4.0","retrieved":"2026-10-12","source":"Epoch AI, 'Capabilities \u0026 benchmarking', epoch.ai/benchmarks"},"modified":"2026-10-12","section":"llm","tags":["llm-benchmarks","leaderboard","model-comparison","ai-progress"],"title":"LLM-Benchmarks 2019–2026: die Bestmarken auf 16 Ranglisten","translations":[{"json":"https://abxk.ai/en/llm/benchmarks/index.json","language":"en","url":"https://abxk.ai/en/llm/benchmarks/"}],"url":"https://abxk.ai/de/llm/benchmarks/"}