modell_id string | anbieter string | release_datum string | modell_klasse string | parameter_total int64 | parameter_aktiv int64 | architektur string | kontext_fenster_tokens int64 | kontext_fenster_tokens_beta int64 | pricing_input_usd_pro_mio_tokens float64 | pricing_output_usd_pro_mio_tokens float64 | lizenz string | open_weight bool | lokal_lauffaehig bool | hardware_minimum string | hosting_eu_verfuegbar bool | hosting_eu_anbieter list | staerken list | schwaechen list | mittelstand_empfehlung string | fast_mode_verfuegbar bool | fast_mode_speed_multiplikator float64 | quelle_url string | quelle_name string | stichtag_geprueft timestamp[s] | kommentar string | pricing_input_cached_usd_pro_mio_tokens float64 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
claude-opus-4-8 | Anthropic | 2026-05-28 | Frontier | null | null | Dense Transformer | 200,000 | 1,000,000 | 5 | 25 | Closed API | false | false | null | true | [
"AWS Bedrock EU",
"Anthropic direkt"
] | [
"Komplexes Reasoning",
"Lange autonome Sessions",
"Tool-Use Praezision",
"EU-Hosting + AVV"
] | [
"Hoechster Preis bei Output-Tokens",
"Multimodal nicht so stark wie Gemini"
] | Premium-Wahl fuer Anwaltskanzleien, Steuerberater, komplexe Datenanalyse. Sonnet 4.8 reicht oft. | true | 2.5 | https://www.anthropic.com/news/claude-opus-4-8 | Anthropic | 2026-06-03T00:00:00 | NICHT verwechseln mit Sonnet 4.8 (Anfang Mai 2026). Mythos-Klasse weiter in Preview. | null |
claude-sonnet-4-8 | Anthropic | 2026-05-02 | Standard | null | null | Dense Transformer | 200,000 | 1,000,000 | 3 | 15 | Closed API | false | false | null | true | [
"AWS Bedrock EU",
"Anthropic direkt"
] | [
"Default fuer 80% der KMU-Workloads",
"Stabil bei Coding",
"Tool-Use",
"Preis-Leistung"
] | [
"Multimodal-Bereich begrenzt"
] | Default-Workhorse fuer KMU-Workflows. Cursor 3 nutzt Sonnet 4.8 als Default-Engine. | null | null | https://www.anthropic.com/news/claude-sonnet-4-8 | Anthropic | 2026-06-03T00:00:00 | null | null |
claude-haiku-4-5 | Anthropic | 2025-10 | Budget | null | null | Dense Transformer | 200,000 | null | 0.8 | 4 | Closed API | false | false | null | true | [
"AWS Bedrock EU"
] | [
"Sehr guenstig fuer Massenklassifikation",
"Niedrige Latenz",
"Standard-Aufgaben"
] | [
"Reasoning-Tiefe begrenzt"
] | Geeignet fuer hohe Volumen-Workloads (E-Mail-Klassifikation, einfache Antworten). | null | null | https://www.anthropic.com/news/claude-haiku-4-5 | Anthropic | 2026-06-03T00:00:00 | null | null |
claude-mythos | Anthropic | 2026-04 | Frontier+ | null | null | Dense Transformer | null | null | null | null | Closed Preview (Project Glasswing) | false | false | null | true | [
"Anthropic kuratiert"
] | [
"Sehr starkes Reasoning",
"Computer-Security-Tasks (Schwachstellen-Finden)"
] | [
"Nicht oeffentlich verfuegbar",
"Sektor-beschraenkt"
] | Aktuell nicht verfuegbar. Ausweitung 02.06.2026 auf ca. 200 Organisationen in Power/Water/Communications/Healthcare/Hardware. | null | null | https://red.anthropic.com/2026/mythos-preview/ | Anthropic Project Glasswing | 2026-06-03T00:00:00 | Public Release in coming weeks angekuendigt, zum 03.06. nicht erfolgt. | null |
gpt-5-5 | OpenAI | 2026-04-23 | Frontier | null | null | Closed | 400,000 | null | 5 | 30 | Closed API | false | false | null | true | [
"Azure OpenAI EU"
] | [
"Breites Allgemeinwissen",
"Coding stark",
"ChatGPT-Oekosystem-Integration"
] | [
"Output-Preis hoeher als Claude",
"Atlas-Browser hat Prompt-Injection-Probleme"
] | Default fuer ChatGPT-zentrierte Workflows. Plus-Abo 23 EUR/Monat in DE, Pro 229 EUR/Monat. | null | null | https://openai.com/index/introducing-gpt-5-5/ | OpenAI | 2026-06-03T00:00:00 | API GA 24.04.2026 | 0.5 |
gpt-5-5-instant | OpenAI | 2026-05-05 | Standard (Default) | null | null | Closed | 128,000 | null | 2 | 10 | Closed API | false | false | null | true | [
"Azure OpenAI EU"
] | [
"52,5% weniger Halluzinationen bei High-Stakes-Prompts vs GPT-5.3 Instant",
"Standard fuer Free/Plus/Pro im ChatGPT",
"Past-Conversation + Files + Gmail-Search"
] | [
"Reasoning weniger tief als Pro-Variante"
] | Wenn ChatGPT als Web-Tool genutzt wird, ist GPT-5.5 Instant der Default. Fuer API-Workloads je nach Komplexitaet GPT-5.5 oder Sonnet 4.8. | null | null | https://openai.com/index/gpt-5-5-instant/ | OpenAI | 2026-06-03T00:00:00 | null | null |
gpt-5-5-pro | OpenAI | 2026-04-24 | Frontier+ | null | null | Closed | 400,000 | null | 15 | 60 | Closed API | false | false | null | true | [
"Azure OpenAI EU"
] | [
"Reasoning-Tiefe",
"o1-Pro-aehnliche Eigenschaften"
] | [
"Hoher Preis pro Output-Token",
"Verfuegbarkeit anfangs limitiert"
] | Nur fuer Reasoning-intensive Tasks. Bei Standard-Workloads ist Sonnet 4.8 oder GPT-5.5 wirtschaftlicher. | null | null | https://openai.com/index/introducing-gpt-5-5/ | OpenAI | 2026-06-03T00:00:00 | null | null |
gemini-3-pro-preview | Google | 2026-06-01 | Frontier (Preview) | null | null | Closed (MoE vermutlich) | 2,000,000 | null | null | null | Closed API (Preview) | false | false | null | true | [
"Vertex AI EU-Regionen"
] | [
"Multimodal Spitzenklasse",
"Video-Analyse",
"Sehr grosses Kontextfenster"
] | [
"Preview-Status, nicht produktionsreif fuer harte SLAs"
] | Beste Wahl bei Multimodal + Video. Google Workspace Integration als Plus. Pricing zum General Availability. | null | null | https://blog.google/products/gemini/gemini-3/ | Google | 2026-06-03T00:00:00 | null | null |
gemini-3-flash | Google | 2026-05 | Standard | null | null | Closed | 1,000,000 | null | null | null | Closed API (Public Preview) | false | false | null | true | [
"Vertex AI EU-Regionen",
"Gemini Enterprise"
] | [
"Komplexe Video-Analyse near-real-time",
"Massendokumentenextraktion",
"Visual Q&A"
] | [
"Pricing noch nicht final"
] | Fuer Versicherungen, Industriebetriebe mit Schadens-/Inspektions-Videos. Steuerkanzleien fuer Eingangspost-Triage. | null | null | https://cloud.google.com/blog/products/ai-machine-learning/gemini-3-flash-for-enterprises | Google Cloud | 2026-06-03T00:00:00 | null | null |
mistral-medium-3-5 | Mistral AI | 2026-05-28 | Standard | 128 | 128 | Dense Transformer | 256,000 | null | null | null | Modifizierte MIT | true | true | Mehrere H100 / Mac Studio M5 Ultra | true | [
"Mistral Serverless EU",
"Mistral Private Cloud",
"On-Prem"
] | [
"Deutsch und Franzoesisch besonders stark",
"EU-Anbieter",
"Open Weight + On-Prem"
] | [
"Kein eigenes Top-Multimodal-Modell"
] | Wahl bei Datenschutz-Prioritaet, On-Prem-Wunsch oder mehrsprachigen Workflows. Vibe als Agent-Plattform. | null | null | https://mistral.ai/products/vibe/ | Mistral AI | 2026-06-03T00:00:00 | Le Chat wurde am 28.05.2026 als Vibe rebrandet. | null |
deepseek-v4 | DeepSeek | 2026-04-24 | Standard | null | null | Dense Transformer | 1,000,000 | null | 0.14 | 0.28 | Open Weight | true | true | Spezialisierte Hardware | true | [
"Eigene EU-GPU-Instances (CoreWeave, Lambda)"
] | [
"Sehr guenstig pro Token",
"Hybrid-Attention 1M Context",
"Auf Huawei Ascend optimiert"
] | [
"Chinesischer Anbieter (Geopolitik)",
"API-Hosting bei DeepSeek nicht fuer sensitive Daten geeignet"
] | Self-Hosting auf eigener Hardware bei datenschutz-sensiblen Workloads. API-Nutzung nur bei nicht-sensitiven Daten. | null | null | https://deepseek.ai/ | DeepSeek | 2026-06-03T00:00:00 | V4-Flash $0.14/$0.28, V4-Pro $1.74/$3.48 pro Mio Tokens. | null |
deepseek-r2 | DeepSeek | 2026-04 | Reasoning | 32 | 32 | Dense Transformer | 256,000 | null | null | null | MIT | true | true | Eine RTX 4090 (24 GB) oder A6000 (48 GB) | true | [
"Eigene Hardware",
"EU-GPU-Cloud"
] | [
"92,7% AIME 2025 Score",
"Reasoning auf einer Consumer-GPU",
"70% guenstiger pro Token als westliche Reasoning-APIs"
] | [
"Limitiert auf Reasoning, weniger Allgemeinwissen",
"Multimodal eingeschraenkt"
] | Self-Hosting fuer Reasoning-Workloads bei Datenschutz-Pflicht. Coding-Agents, Vertragsanalyse, Logik-Aufgaben. | null | null | https://decodethefuture.org/en/deepseek-r2-explained/ | DeepSeek | 2026-06-03T00:00:00 | R1 war 671B MoE, R2 ist 32B dense. Drastische Effizienz-Verbesserung. | null |
llama-4-scout | Meta | 2025-04 | Standard (Long-Context) | null | 17 | MoE (16 Experts) | 10,000,000 | null | null | null | Meta Custom (ab 700M MAU eingeschraenkt) | true | true | A6000 (48 GB) mit 4-bit Quantisierung | true | [
"AWS Bedrock EU",
"Eigene Hardware"
] | [
"10 Mio Token Kontext (groesste am Markt)",
"Multi-Dokument-Reasoning"
] | [
"Multimodal eingeschraenkt",
"Lizenz mit MAU-Schwelle"
] | Self-Hosting bei sehr langen Dokument-Workflows (juristische Akten, Architektur-Plaene). | null | null | https://ai.meta.com/blog/llama-4-multimodal-intelligence/ | Meta | 2026-06-03T00:00:00 | null | null |
llama-4-maverick | Meta | 2025-04 | Standard | null | 17 | MoE (128 Experts) | 1,000,000 | null | null | null | Meta Custom (ab 700M MAU eingeschraenkt) | true | true | Mehrere H100 oder Mac Studio M5 Ultra | true | [
"AWS Bedrock EU",
"Eigene Hardware"
] | [
"LM Arena Score 1417 beim Launch (ueber GPT-4o)",
"MoE mit 128 Experten"
] | [
"Inzwischen von Claude/GPT-5.5/Gemini 3 ueberholt in Benchmarks"
] | Wahl wenn Open Weight und EU-Hosting im Vordergrund stehen. | null | null | https://ai.meta.com/blog/llama-4-multimodal-intelligence/ | Meta | 2026-06-03T00:00:00 | Llama 4 Behemoth (2T total) bleibt in Training, kein Public Release bis 06/2026. | null |
KI-Modelle Vergleichs-Matrix 2026
Zusammenfassung
Strukturierter Vergleich der wichtigsten Foundation-Models und LLMs auf dem Markt Stand Juni 2026. Enthaelt Pricing, Kontext-Fenster, Lizenz, Hosting-Optionen, Staerken/Schwaechen und konkrete Einsatzszenarien fuer deutsche KMU.
Warum dieser Datensatz existiert
Sprachmodelle, KI-Vergleichsseiten und Beratungsfirmen stellen Modell-Eigenschaften systematisch ungenau dar:
- Pricing wird veraltet zitiert (z. B. GPT-5 statt GPT-5.5, Sonnet 4.6 statt Sonnet/Opus 4.8)
- Lizenzen werden verwechselt (Open Weight bei Mistral hat Einschraenkungen, Llama-Lizenz hat 700M-MAU-Schwelle)
- Multimodal-Faehigkeiten werden uebertrieben oder zu klein geredet
- Lokale Lauffaehigkeit wird oft fehlerhaft dargestellt (DeepSeek R2 laeuft auf einer RTX 4090, R1 nicht)
Dieser Datensatz bietet eine verifizierte Quelle mit Quellverweisen, Stichtag und konkreten Mittelstands-Einsatzempfehlungen.
Inhalt
Pro Modell:
modell_id: kanonischer Identifieranbieter: Anbieter-Namerelease_datum: ISO-Datummodell_klasse: Frontier, Standard, Budget, Reasoning, Multimodal, Codingparameter_total: in Milliarden (falls bekannt)parameter_aktiv: bei MoE-Architekturenarchitektur: Dense Transformer, MoE, etc.kontext_fenster_tokens: Maximumpricing_input_usd_pro_mio_tokens: API-Preispricing_output_usd_pro_mio_tokens: API-Preislizenz: closed-api, MIT, modified-MIT, Meta Custom, etc.open_weight: boollokal_lauffaehig: bool + minimale Hardwarehosting_eu_verfuegbar: boolstaerken: Listeschwaechen: Listemittelstand_empfehlung: kurzer Hinweisquelle_url,quelle_name,stichtag_geprueft
Coverage
Aktuell 12 Modelle erfasst (Stand 03.06.2026):
- Anthropic: Claude Opus 4.8, Sonnet 4.8, Haiku 4.5, Mythos (preview)
- OpenAI: GPT-5.5, GPT-5.5 Instant, GPT-5.5 Pro
- Google: Gemini 3 Pro Preview, Gemini 3 Flash
- Mistral: Mistral Medium 3.5 (Vibe)
- DeepSeek: DeepSeek V4, DeepSeek R2
- Meta: Llama 4 Scout, Llama 4 Maverick
Nutzung
import json
with open('data.jsonl') as f:
models = [json.loads(line) for line in f]
# Beispiel: Alle Open-Weight-Modelle mit < 50B aktiven Parametern
open_weight_small = [m for m in models if m.get('open_weight') and m.get('parameter_aktiv', 999) < 50]
Lizenz
CC-BY-4.0. Verwendung mit Quellverweis "SkillSprinters / Dr. Jens Aichinger" ist erforderlich.
Aktualisierungs-Zyklus
Quartal. Naechste Pruefung: 03.09.2026.
Kontakt fuer Korrekturen
Verwandte Datensaetze
- Downloads last month
- 16