Why these rules

    Every tool in the index is classified with the same criteria. Here they are, in full — so you can disagree with them on the record.

    Jurisdiction tiers

    EU / EEA

    EU member state or EEA (Norway, Iceland, Liechtenstein) — GDPR applies directly.

    Adequacy country

    Non-EU country with an EU adequacy decision — data flows freely; note the caveat.

    Outside scope

    No adequacy and/or foreign-government exposure — excluded (US CLOUD Act, China) or 'watch' (Ukraine).

    Badges

    EU company

    Vendor headquartered in Europe (state country + tier).

    EU data residency

    EU data-residency / processing option available.

    Open weights

    Open weights / self-hostable.

    Compliance rubric

    Derived from the verified classification data on this site: Very high = EU/EEA jurisdiction + documented EU data residency + self-hostable option. High = EU/EEA + EU data residency. Medium = adequacy country + residency, or EU/EEA without documented residency. Each underlying claim carries its own source.

    AI Capability Index

    Composite index normalised so the highest-scoring model on the Artificial Analysis Intelligence Index = 100. Coding subscore normalised the same way from SWE-bench Verified; where a score is vendor-reported rather than independently run, the source is labelled accordingly. Note that SWE-bench Verified is approaching saturation at the frontier and several vendors no longer publish it, so the coding subscore is a floor, not a ranking. Scores are updated manually with each monthly review and every number links to its source and retrieval date. Models without published independent benchmarks show 'Not yet scored', with a stated reason. Where a tool publishes benchmarks that are not comparable to the two normalised scores (for example HumanEval, image-arena Elo or word error rate), those are listed separately as published benchmarks.

    Primary sources

    Planned sources: LMArena, LiveBench.

    Scores retrieved 2026-09-04. Normalisation anchor: model: Claude Fable 5.1 (max effort) · aaIndex: 66 · sweBenchVerified: 96 · note: Anchor moved from Claude Opus 5 (AA 63) to Claude Fable 5.1 (AA 66) at the September 2026 review; all normalised scores were recalculated..

    Reference models (not listed in the directory)

    Claude Fable 5.1 — Anthropic (US)

    Capability index: 100 · AA Index: 66

    Reference only — non-EU, not listed in the directory.

    Claude Opus 5 — Anthropic (US)

    Capability index: 95 · Coding: 100 · AA Index: 63 · SWE-bench Verified: 96

    Reference only — non-EU, not listed in the directory.

    GPT-5.6 Sol — OpenAI (US)

    Capability index: 92 · AA Index: 61

    Reference only — non-EU, not listed in the directory.

    Gemini 3.7 Flash — Google (US)

    Capability index: 85 · AA Index: 56

    Reference only — non-EU, not listed in the directory.

    Review cadence and disclaimer

    Maintained by Stefano Vincenti / aitrainer.dk. Last full review 2026-09-04, reviewed monthly.

    Independent, informational guide — not legal, security or compliance advice. Classifications describe company location, GDPR transfer status, data-residency options and licensing as of the date shown, from public sources. Verify data-processing terms with the provider before use, especially for personal or sensitive data. No affiliation with any vendor.