Lexiel vs alternatives: beyond the percentage
Accuracy matters. But so does how the attorney works with the tool.
| Lexiel RAG jurídico | Claude Opus 4.6 sin RAG legal | Gemini 3.1 Pro sin RAG legal | GPT-5.4 thinking · sin RAG | |
|---|---|---|---|---|
| Benchmark accuracy 2026 | 98.1% 304 out of 311 questions | 94% measured | 95,3% measured | 84,7% measured |
| Updated BOE corpus | ✓ 70 full laws | ✗ | ✗ | ✗ |
| Supreme + Constitutional Court | ✓ 1.868 sentencias | ✗ | ✗ | ✗ |
| Exact article citation | ✓ art. X, párr. N | ∼ | ∼ | ∼ |
| Per-case history | ✓ | ✗ | ✗ | ✗ |
| Word / Office plug-in | ✓ | ✗ | ✗ | ✗ |
| Data processed in Spain (GDPR) | ✓ | ✗ | ✗ | ✗ |
| Price per attorney/month | €69 - €99 | API / ~€20 | Free / €22 | ~€24 |
∼ = partial support. Claude Opus 4.6, Gemini 3.1 Pro and GPT-5.4 scores measured directly on the same 150-question test without specialized legal corpus : Lexiel with full legal RAG. Indicative pricing Q1 2026.
Legal accuracy comparison
Bar access exam · 150 questions · 2023/2024/2025
* Generic AIs tested without access to specialized legal corpus. Scores reproducible with official Ministry of Justice answer keys.
Where does the advantage come from?
The difference is not the base model. It's the verified legal corpus.
Each answer is complemented with exact BOE, CENDOJ and Constitutional Court fragments retrieved by semantic similarity.
Best model without specialized corpus: Gemini 3.1 Pro (95.3%). We also tested Claude Opus 4.6 (94%) and GPT-5.4 with thinking (84.7%). They respond from general training, without citing exact articles or rulings.
The legal corpus advantage is structural: BOE + CENDOJ + 1,868 Supreme/Constitutional Court rulings allow citing the exact norm, not paraphrasing it.
Breakdown by area of law
All areas above 85%
Benchmark evolution
Every corpus update improves accuracy
+8 points in 4 months; each ingestion of new BOE/CENDOJ sources improves thematic coverage and reduces errors from missing regulatory context.
Why are there two published figures?
In some blog posts you'll find 98.3%. Here we publish 99.3%. It's not an error; they are two different tests on the same official exam.
Test A: 100 questions (2024+2025)
98,3%
Used for direct comparison with Maite AI, which published results on these same sessions.
Test B: 150 questions (2023+2024+2025) ★ official
99,3%
Three sessions, more jurisdictional competence questions (LOPJ). Harder and more representative. This is the official figure.
What we got wrong and why
3 errors out of 156 (official test). Here it is, with cause and solution in progress.
Question about LOPJ Arts. 38-42 (inter-jurisdictional conflicts). The model retrieves the correct provisions but picks the wrong option on a very specific civil vs. administrative jurisdiction conflict scenario.
None of the errors are fabricated hallucinations. 100% of failures are incorrect disambiguations of real norms or reforms not yet incorporated into the corpus. This is Lexiel's commitment: traceable errors, not fabricated answers.
"I evaluated the answers in completely blind mode, without knowing which tool generated each response. Lexiel was the only one that consistently included the exact article citation and ruling number, not a paraphrase. In 20 years of practice, that's what differentiates a professional brief from an amateur one."
Sample verified answers
Every answer includes the exact article or case citation.
En el contrato de compraventa, ¿cuándo se entiende transmitida la propiedad según el Código Civil español?
Con la traditio o entrega efectiva de la cosa vendida (art. 609 y 1462 CC), no con el mero consentimiento (a diferencia del sistema francés).
Art. 609, 1445 y 1462 Código Civil
Explique la diferencia entre dolo eventual y culpa consciente con imprudencia grave.
En el dolo eventual el sujeto prevé el resultado como probable y lo acepta (teoría del consentimiento, STS 26/11/2008). En culpa consciente, lo prevé pero confía en que no se producirá.
Art. 10-12 CP; STS 26/11/2008 (ECLI:ES:TS:2008:6507)
¿Qué recurso cabe contra una ley autonómica que vulnera el bloque de constitucionalidad?
Recurso de inconstitucionalidad (art. 161.1.a CE) ante el TC, legitimados el Presidente del Gobierno, el Defensor del Pueblo, 50 Diputados, 50 Senadores, gobiernos y parlamentos autonómicos. Plazo: 3 meses (art. 33 LOTC).
Art. 161-163 CE; Art. 31-34 LOTC
En el concurso de acreedores, ¿qué es la masa activa concursal?
El conjunto de bienes y derechos del deudor integrados en el concurso para satisfacer a los acreedores (art. 192 TRLC), incluyendo los adquiridos durante el concurso y excluyendo los inembargables por ley.
Art. 192-200 Texto Refundido de la Ley Concursal (TRLC)
Frequently asked questions
What every attorney asks before trusting a legal AI.
The methodology is public and reproducible. Use the same official questions and share your results. We will update the comparison table with any verified result.
Methodology
Reproducible, transparent, third-party audited.
150 official questions
Extracted from the official bar access exams 2023, 2024 & 2025 published by Spain's Ministry of Justice (BOE).
Verified answers
Correct answers extracted from the Ministry of Justice's official answer keys.
RAG pipeline active
Lexiel uses its verified legal corpus (BOE, CENDOJ), no external internet access.
Reproducible score
Test run with official public questions and answers. Anyone can verify the results.
Question source
Scoring rules & human context
The bar access exam uses penalized scoring: each wrong answer deducts 0.25 points. Scoring 144/145 correct is nearly perfect, far above the required minimum.
Download the benchmark logs
Each test exports a JSON with question, model's chosen option, correct answer and whether RAG was used. Replicate the benchmark with the same official Ministry of Justice questions.
Format: JSON with results[] field, each entry includes id, question, correct answer, model answer, isCorrect, hadRag. Questions are public (BOE). You can reproduce the test on any LLM and compare.
Spanish Judiciary Access Exam (CGPJ)
The most demanding competitive exam in the Spanish judicial system. Lexiel is evaluated against the same official CGPJ questions that candidates face.
99 out of 100 questions correct
Gemini 3.1 Pro + RAG · Mar 2026
190 out of 200 questions correct
Base model without specialized corpus
Judiciary Methodology
Reproducible, transparent, verified CGPJ sources.
257 official CGPJ questions
Multiple-choice questions extracted from the official Judiciary Competition 2023 and 2025 sittings, published by Spain's General Council of the Judiciary.
Verified official answer keys
Correct answers cross-checked against the official CGPJ answer keys published on their judicial competition portal.
RAG pipeline over CGPJ + BOE corpus
Corpus includes TS, AN, TC case law and state legislation indexed from BOE and CENDOJ, without external internet access.
Reproducible scoring
Questions and answer keys are public CGPJ documents. Any competitor or external auditor can replicate the test with the same sources.
Question source
Scoring rules & exam context
The Judiciary Competition is one of the most competitive in the Spanish legal system: limited seats (~100-200/year), thousands of candidates, and very high cut-off scores. The knowledge test is only one phase of the full evaluation.
Multi-country legal corpus
Lexiel indexes case law and legislation from 20 jurisdictions. 130,000+ official sources, 2,590,000+ verified provisions.
Get the May 2026 benchmark
Each new edition in your inbox. Results, corpus changes and model improvements.
No spam. Unsubscribe anytime.