xi6eal/ki-benchmark-deutsch

i6eal KI-Benchmark Deutsch — German Business AI Benchmark

Results of a recurring benchmark that measures leading LLMs on real German business tasks — one of the first leaderboards built specifically for the German language and German domain knowledge rather than translated English test sets.

Live leaderboard: https://i6eal.de/ki-benchmark-deutsch/ Archived on Zenodo: doi:10.5281/zenodo.21452970 This release: benchmark run of 2026-06-22 · 7 models · 6 categories · 24 tasks per model · scale 0–100

Key findings (run 2026-06-22)

• Claude Opus 4.8 leads overall with 98/100, ahead of GPT-5.5 (95) and Qwen3.7 Max (93).
• Amts- & Behördendeutsch (official/bureaucratic German) is the weakest category for every model family — category average 83.4, versus 97.1 for source-grounded answering (RAG) and 96.3 for language quality. Gemini 3.1 Pro (79), Grok 4.3 (78) and Mistral Large 3 (79) all drop hardest here.
• Efficient-class models are closer than expected: Qwen3.7 Max (93) and DeepSeek V4-Pro (91) match or beat flagship Gemini 3.1 Pro (91) and Grok 4.3 (90) overall.
• Verifiable law & tax questions (§§ BGB/AO, deadlines, limitation periods) split the field: 100 (GPT-5.5) and 99 (Opus 4.8) down to 74–77 (DeepSeek V4-Pro, Grok 4.3) — hallucinated section numbers score as wrong.

Categories

ID	Category (DE)	What it tests	Scoring
business	Geschäftskommunikation	Formal business correspondence — emails, quotes, rejections in the right register	judge
amtsdeutsch	Amts- & Behördendeutsch	Understanding bureaucratic German and rendering it in plain language	judge+reference
recht	Recht & Steuern	Verifiable questions on German law and tax (BGB, AO, deadlines)	reference
zusammenfassung	Zusammenfassung & Treue	Faithful summarisation — figures unchanged, nothing fabricated	judge+reference
rag	Quellentreue (RAG)	Answering strictly from a provided document, refusing when absent	judge+reference
sprache	Sprachqualität	Rewriting clumsy text into idiomatic business German	judge

Full rubrics (DE/EN) are in categories.csv.

Files

File	Contents
results.csv	One row per model × category score (42 rows)
models.csv	Model-level metadata and overall scores
categories.csv	Category definitions and scoring rubrics, DE + EN
public_samples.jsonl	One public example prompt per category
snapshot.json	Full raw snapshot of this run as published on i6eal.de

Methodology

• Judging: subjective categories are scored by a cross-vendor panel of neutral models (OpenAI GPT-5.4, Google Gemini 3 Pro, Anthropic Claude Opus 4.7) that are not themselves on the leaderboard; leave-one-family-out, name-blind, spot-checked by hand. Reference categories are scored against the correct answer; hallucinated statutes or deadlines count as wrong.
• Held-out test set: only the sample prompts in this dataset are public. The actual test set stays private so the benchmark cannot be trained on or gamed. This dataset therefore publishes results, rubrics and samples, not the test items themselves.
• Cadence: the benchmark runs regularly; every run is versioned. New runs will be published as new versions of this dataset.

Limitations

• 7 models per run; the field moves quickly and older runs describe older model versions.
• Judge-based scores are model-panel judgments against a fixed rubric, not human expert ratings (though spot-checked by hand).
• 24 tasks per model per run — designed for stable rankings and category diagnostics, not for fine-grained sub-point differences between adjacent scores.

Citation

@dataset{syka_ki_benchmark_deutsch_2026,
  author    = {Syka, Ideal},
  title     = {i6eal KI-Benchmark Deutsch — German Business AI Benchmark (run 2026-06-22)},
  year      = {2026},
  publisher = {i6eal.de / Syka Ventures UG (haftungsbeschränkt)},
  url       = {https://i6eal.de/ki-benchmark-deutsch/}
}

License

CC BY 4.0 — free to use, share and adapt with attribution to i6eal.de. Suggested attribution: „Quelle: i6eal KI-Benchmark Deutsch (i6eal.de)".

Contact: ai@i6eal.de