Ranking IA
The ranking of AI models in Spanish
11 Frontier models evaluated
7 Labs
1,100 Scored responses
02 / 06
Background
A benchmark built in Spanish, with every card on the table
Most AI model rankings are measured in English, and not every model performs the same when the work is in Spanish. Ranking IA compares frontier models with 25 real-world tests written in Spanish and grouped into six families: instructions, truthfulness, reasoning, character, expression and Spanish.
Every test is run four times on every model. Whatever a script can check is scored by a verifier; the rest goes to a panel of judges from four labs, and none of them grades the models of its own lab. The methodology is published before anything runs, and the author scores nothing.
The result is a tier table, from S to D, with the overall score, the 95% confidence interval and the price of each model. No cell is left empty: every one opens to show the original response, and the full data can be downloaded as JSON under a CC BY 4.0 license.
The ranking is updated with every new measurement, and each model keeps the run in which it was measured, so the comparison stays fair.
03 / 06
Results
25
Real-world tests
6
Families evaluated
4
Runs per test
4
Judges with recusal by lab
04 / 06
Scope
A Sozpic LAB product
2.1
Services
- Methodology design and pre-registration
- A bank of 25 real-world tests in Spanish
- Panel of automated judges with recusal by lab
- Bootstrap confidence intervals
- Public website with tiers, a browsable table and head-to-head comparison
- Open data, published in Spanish and English
2.2
Technologies
- Models from Anthropic, OpenAI, Google, xAI, DeepSeek, Moonshot and Mistral
- Script verifiers and LLM judges
- 1,000-iteration bootstrap for the 95% confidence interval
- Full data as JSON under a CC BY 4.0 license
