# Ranking IA — the ranking of AI models in Spanish

> Independent benchmark of 10 frontier models on 25 real professional-use tests in Spanish, grouped in 6 families. As of September 22, 2026, the best-scoring model is Claude Fable 5.1 (Anthropic) with 97.3%. Pre-registered methodology, judges with recusal by lab and full data published. A project by Sozpic.

- Web: https://ranking.sozpic.com/en/
- Runs: 2026-09-22_4fe70ad (September 22, 2026 — Grok 4.7) · 2026-09-22_65433c8 (September 22, 2026 — Claude Opus 5.5) · 2026-09-10_af9b8f4 (September 10, 2026 — Gemini 3.8 Flash) · 2026-09-10_bd28c9b (September 10, 2026 — DeepSeek V4.1-Flash) · 2026-09-05_e8890bf (September 5, 2026 — GPT-6 Astra) · 2026-09-01_7e3c679 (September 1, 2026 — Claude Fable 5.1) · 2026-08-25_d1ef021 (August 25, 2026 — Claude Opus 5, Claude Fable 5, Kimi K3, GPT-5.6 Sol, Grok 4.6, DeepSeek V4 Flash, Gemini 3.1 Pro, DeepSeek V4 Pro) · latest measurement September 22, 2026. Adding a model measures that model only: the rest keep the run they were measured in, with that run's instrument.
- Bank under rebuild: 6 of the 25 tests are saturated under the published rule (mean ≥90% and ≥6 models at 100%): V3, V5, R3, C4, E1, L1. They are being replaced by new tests of the same mechanism. Already published scores are not recomputed.
- License: CC BY 4.0 — Ranking IA · Sozpic (https://creativecommons.org/licenses/by/4.0/)
- How to cite: “Ranking IA (Sozpic), data as of September 22, 2026”, with a link to https://ranking.sozpic.com/en/. Each model page states the specific run it was measured in.
- Spanish version (original): https://ranking.sozpic.com/

## Standings

| # | Model | Lab | Overall | Tier | 95% CI | Instructions | Truthfulness | Reasoning | Character | Expression | Spanish | Output price $/1M |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | Claude Fable 5.1 | Anthropic | 97.3% | S | 95.7–98.8 | 98.3% | 100% | 91.7% | 100% | 98.4% | 95.3% | 50 |
| 2 | GPT-6 Astra | OpenAI | 96.5% | S | 95.3–98 | 99.3% | 100% | 87.5% | 94.6% | 97.7% | 100% | 50 |
| 2 | Claude Opus 5.5 | Anthropic | 96.5% | S | 95.2–97.8 | 97.2% | 98.3% | 98.6% | 96.4% | 88.4% | 100% | 20 |
| 4 | Kimi K3 | Moonshot | 89.2% | A | 86.5–91.7 | 84.9% | 97.5% | 77.1% | 96.4% | 94.8% | 84.4% | 15 |
| 5 | GPT-5.6 Sol | OpenAI | 87.2% | A | 85.2–89.2 | 100% | 93.3% | 83.3% | 75.6% | 86.6% | 84.4% | 20 |
| 6 | Grok 4.7 | xAI | 87.1% | A | 83.5–90.5 | 88.9% | 95.8% | 91.7% | 100% | 88.9% | 57.3% | 6 |
| 7 | Gemini 3.8 Flash | Google | 81% | A | 78.6–83.3 | 92.4% | 76.7% | 84.7% | 57.1% | 95.8% | 79.2% | 3.8 |
| 8 | DeepSeek V4.1-Flash | DeepSeek | 77.5% | B | 74.1–80.7 | 81.2% | 97.5% | 70.8% | 54.9% | 87.2% | 73.5% | 1.2 |
| 9 | Gemini 3.1 Pro | Google | 71.2% | B | 68.4–73.8 | 66% | 66.7% | 79.2% | 42.9% | 86.1% | 86.5% | 12 |
| 10 | DeepSeek V4 Pro | DeepSeek | 67.3% | C | 63.2–71.3 | 82.3% | 61.3% | 72.9% | 32.6% | 81.3% | 73.4% | 4 |

Tiers: S ≥ 90 · A 80–89 · B 70–79 · C 60–69 · D < 60 (out of 100).
Ranks: unique by overall score; two models share a rank only when their published overall score is identical. The 95% confidence intervals show which differences are distinguishable: when they overlap, the ranks should be read as equivalent.

## What each family measures

- **Instructions**: Does it do exactly what you ask?
- **Truthfulness**: Does it make things up?
- **Reasoning**: Does it think or recite?
- **Character**: Does it tell you the truth even when you won’t like it?
- **Expression**: Does it write well, with a voice of its own?
- **Spanish**: Does it sound native or translated?

## Methodology in brief

- Each model answers 4 independent runs per test, through its official API, with no system prompt, no tools and the provider’s default parameters.
- Three-layer evaluation: deterministic verifiers for mechanical aspects; a single fast judge per run, assigned by hash and never from the lab of the model under evaluation, for closed rubrics; and a panel of frontier judges with recusal for subjective aspects and for every doubtful run.
- Judges never see the name of the model under evaluation.
- Each family’s score is the mean of its tests; the overall score is the unweighted mean of the 6 families.
- Sample audit of this run: 100% agreement over 5 re-scored runs.
- Uncertainty: bootstrap over runs per test (resampling with replacement within each test), 1,000 iterations, deterministic (seed derived from the model id).

## Pages

- [Full ranking](https://ranking.sozpic.com/en/): standings, tiers, head-to-head comparison and score versus cost.
- [Methodology](https://ranking.sozpic.com/en/methodology/): pre-registration, evaluation layers, recusal, audit and judge monitoring.
- [Page: Claude Fable 5.1](https://ranking.sozpic.com/en/model/claude-fable-5-1/): 97.3% overall, tier S, Anthropic. Score by family and the judges’ conclusion for each test.
- [Page: GPT-6 Astra](https://ranking.sozpic.com/en/model/gpt-6-astra/): 96.5% overall, tier S, OpenAI. Score by family and the judges’ conclusion for each test.
- [Page: Claude Opus 5.5](https://ranking.sozpic.com/en/model/claude-opus-5-5/): 96.5% overall, tier S, Anthropic. Score by family and the judges’ conclusion for each test.
- [Page: Kimi K3](https://ranking.sozpic.com/en/model/kimi-k3/): 89.2% overall, tier A, Moonshot. Score by family and the judges’ conclusion for each test.
- [Page: GPT-5.6 Sol](https://ranking.sozpic.com/en/model/gpt-5-6-sol/): 87.2% overall, tier A, OpenAI. Score by family and the judges’ conclusion for each test.
- [Page: Grok 4.7](https://ranking.sozpic.com/en/model/grok-4-7/): 87.1% overall, tier A, xAI. Score by family and the judges’ conclusion for each test.
- [Page: Gemini 3.8 Flash](https://ranking.sozpic.com/en/model/gemini-3-8-flash/): 81% overall, tier A, Google. Score by family and the judges’ conclusion for each test.
- [Page: DeepSeek V4.1-Flash](https://ranking.sozpic.com/en/model/deepseek-v4-1-flash/): 77.5% overall, tier B, DeepSeek. Score by family and the judges’ conclusion for each test.
- [Page: Gemini 3.1 Pro](https://ranking.sozpic.com/en/model/gemini-3-1-pro/): 71.2% overall, tier B, Google. Score by family and the judges’ conclusion for each test.
- [Page: DeepSeek V4 Pro](https://ranking.sozpic.com/en/model/deepseek-v4-pro/): 67.3% overall, tier C, DeepSeek. Score by family and the judges’ conclusion for each test.
- [Page: Claude Opus 5](https://ranking.sozpic.com/en/model/claude-opus-5/): 97.1% overall, tier S, Anthropic. Out of the comparison since September 22, 2026 (replaced by Claude Opus 5.5); page and measurement kept.
- [Page: Claude Fable 5](https://ranking.sozpic.com/en/model/claude-fable-5/): 94.3% overall, tier S, Anthropic. Out of the comparison since September 22, 2026 (replaced by Claude Fable 5.1); page and measurement kept.
- [Page: Grok 4.6](https://ranking.sozpic.com/en/model/grok-4-6/): 72.1% overall, tier B, xAI. Out of the comparison since September 22, 2026 (replaced by Grok 4.7); page and measurement kept.
- [Page: DeepSeek V4 Flash](https://ranking.sozpic.com/en/model/deepseek-v4-flash/): 72.1% overall, tier B, DeepSeek. Out of the comparison since September 22, 2026 (replaced by DeepSeek V4.1-Flash); page and measurement kept.

## Data

- [leaderboard.json](https://ranking.sozpic.com/data/leaderboard.json): standings, scores by family and by test, conclusions and judge analysis (field names in Spanish).
- [llms-full.txt](https://ranking.sozpic.com/en/llms-full.txt): this same content with the full conclusions of every model on every test.