SOTA Benchmark MatrixRohwerte pro Quelle, stabile Modellslots, Version steht als Zell-Subtext.Generiert: 2026-06-07 17:14 UTC
AA500 Modelle Arena200 Modelle SWE40 Einträge Slots bleiben stabil, Werte aktualisieren sich pro Quelle.
ModellslotZugriffAA IQAA CodingAA AgenticGPQA %MMMU-Pro %Arena EloSWE VerifiedSWE DatumFlags
Claude OpusAnthropicAPI61.4Claude Opus 4.8 (max)56.7Claude Opus 4.8 (max)77.8Claude Opus 4.8 (max)92.0Claude Opus 4.8 (max)1504claude-opus-4-6-thinking75.6%378/500 · Claude Opus 4.62026-02-17SWE-benchctx 1000k
Claude SonnetAnthropicAPI51.7Claude Sonnet 4.6 (max)50.9Claude Sonnet 4.6 (max)63.0Claude Sonnet 4.6 (max)87.5Claude Sonnet 4.6 (max)73.3Claude Sonnet 4.6 (max)1470claude-sonnet-4-671.4%357/500 · Claude 4.5 Sonnet (high reasoning)2026-02-17SWE-benchctx 1000k
GPT FlagshipOpenAIAPI60.2GPT-5.5 (xhigh)59.1GPT-5.5 (xhigh)74.1GPT-5.5 (xhigh)93.5GPT-5.5 (xhigh)79.9GPT-5.5 (xhigh)1482gpt-5.5-high72.8%364/500 · GPT-5-2 (high reasoning)2026-02-17SWE-benchctx 922k
Gemini Pro/FlashGoogleAPI57.2Gemini 3.1 Pro Preview55.5Gemini 3.1 Pro Preview59.1Gemini 3.1 Pro Preview94.1Gemini 3.1 Pro Preview82.4Gemini 3.1 Pro Preview1488gemini-3.1-pro-preview75.8%379/500 · Gemini 3 Flash (high reasoning)2026-02-17SWE-benchctx 1000k
Qwen Max/CoderAlibabaAPI/Weights56.6Qwen3.7 Max50.1Qwen3.7 Max66.6Qwen3.7 Max92.3Qwen3.7 Max1474qwen3.7-max-preview55.4%277/500 · Qwen3-Coder 480B/A35B Instruct2025-08-02SWE-benchctx 1000k · SWE open model
DeepSeek ReasonerDeepSeekAPI/Weights51.5DeepSeek V4 Pro (Max)47.5DeepSeek V4 Pro (Max)67.2DeepSeek V4 Pro (Max)88.8DeepSeek V4 Pro (Max)1458deepseek-v4-pro-thinking70.0%350/500 · DeepSeek V3.2 (high reasoning)2026-02-17SWE-benchopen · ctx 1000k
Llama InstructMeta/NVIDIAWeights18.7Llama Nemotron Super 49B v1.515.1Llama Nemotron Super 49B v1.59.4Llama Nemotron Super 49B v1.574.8Llama Nemotron Super 49B v1.51347llama-3.1-nemotron-ultra-253b-v121.0%105/500 · Llama 4 Maverick Instruct2025-07-20SWE-benchopen · ctx 128k
Mistral LargeMistralAPI/Weights22.8Mistral Large 322.7Mistral Large 321.7Mistral Large 368.0Mistral Large 355.7Mistral Large 31415mistral-large-3open · ctx 256k
GemmaGoogleWeights39.2Gemma 4 31B38.7Gemma 4 31B40.9Gemma 4 31B85.7Gemma 4 31B73.4Gemma 4 31B1451gemma-4-31bopen · ctx 256k
Consumer LocalMacBook/Ollama/MLXLocalKein Modellslot: lokale Deployment-Klasse. Sauberer Vergleich braucht konkretes GGUF/MLX, Quantisierung, Hardware, Kontextlänge und Messskript.
Quellen: Artificial Analysis, LMArena, SWE-bench. Rohwerte je Quelle. Unterschiedliche getestete Versionen stehen direkt in der Zelle. Methodik: keine Aggregation, keine Normalisierung, kein Ranking über Quellen hinweg.