Tabla de benchmarks LLM
Puntuaciones de inteligencia, programación y agente lado a lado, con valor por 1M tokens.
Datos del 2026-09-30103 modelos puntuadosArtificial Analysis
| Puesto | ||||
|---|---|---|---|---|
| 1 | Claude Opus 5.5 Anthropic · Adaptive Reasoning, Max Effort, Default Fallback | 57.6 | 6.5 | $4.40 / $22.00 |
| 2 | Claude Sonnet 5.5 Anthropic · Adaptive Reasoning, Max Effort, Default Fallback | 56.0 | 14 | $2.00 / $10.00 |
| 3 | Claude Fable 5.1 Anthropic · Adaptive Reasoning, Max Effort, Default Fallback | 53.4 | 5.3 | $5.00 / $25.00 |
| 4 | Qwen3.8 Max qwen | 53.4 | — | — |
| 5 | GPT-6 Astra OpenAI · max | 52.7 | 0.4 | $60.00 / $300 |
| 6 | GPT-6.1 Sol OpenAI · max | 51.8 | 12 | $2.20 / $11.00 |
| 7 | Claude Opus 5 Anthropic · Adaptive Reasoning, Max Effort | 50.8 | 4.6 | $5.50 / $27.50 |
| 8 | Claude Fable 5 Anthropic · Adaptive Reasoning, Max Effort, Opus 4.8 Fallback | 49.6 | 2.5 | $10.00 / $50.00 |
| 9 | GPT-6 Sol OpenAI · max | 47.5 | 12 | $2.00 / $10.00 |
| 10 | GPT-5.6 Sol OpenAI · max | 47.0 | 5.9 | $4.00 / $20.00 |
| 11 | Grok 4.7 SpaceXAI · xhigh | 46.4 | 14 | $2.20 / $6.60 |
| 12 | MiMo-V2.6-Pro Xiaomi | 46.3 | 85 | $0.435 / $0.870 |
| 13 | Qwen3.8 Max Qwen · 0902 | 45.4 | 15 | $2.00 / $6.00 |
| 14 | GLM-5.3 Z.ai · max | 44.8 | 19 | $1.54 / $4.84 |
| 15 | Grok 4.6 SpaceXAI · high | 44.3 | 13 | $2.20 / $6.60 |
| 16 | Kimi K3 MoonshotAI · max | 43.6 | 7.8 | $2.80 / $14.00 |
| 17 | GPT-5.6 Terra OpenAI · max | 42.1 | 4.7 | $4.00 / $24.00 |
| 18 | Claude Opus 4.8 Anthropic · Adaptive Reasoning, Max Effort | 41.8 | 2.1 | $10.00 / $50.00 |
| 19 | GLM 5.3 Flash Z.ai | 41.8 | 176 | $0.150 / $0.500 |
| 20 | Gemini 3.8 Flash Google · high | 40.9 | 15 | $1.35 / $6.75 |
| 21 | Qwen3.8 2.4T A95B Qwen | 39.9 | 13 | $2.00 / $6.00 |
| 22 | Muse Spark 1.2 Meta · xhigh | 39.6 | 20 | $1.25 / $4.25 |
| 23 | DeepSeek V4.1 Flash DeepSeek · Reasoning, Max Effort | 39.5 | 50 | $0.450 / $1.80 |
| 24 | Gemini 3.7 Flash Google · high | 39.1 | 14 | $1.35 / $6.75 |
| 25 | Grok 4.5 SpaceXAI · high | 38.8 | 6.5 | $4.00 / $12.00 |
| 26 | GPT-5.5 OpenAI · xhigh | 38.4 | 6.8 | $2.50 / $15.00 |
| 27 | Claude Sonnet 5 Anthropic · Adaptive Reasoning, Max Effort | 38.2 | 8.7 | $2.20 / $11.00 |
| 28 | MiMo-V2.6-Flash Xiaomi | 37.9 | 217 | $0.140 / $0.280 |
| 29 | GPT-6 Luna OpenAI · max | 37.3 | 186 | $0.100 / $0.500 |
| 30 | GPT-5.6 Luna OpenAI · max | 37.3 | 41 | $0.400 / $2.40 |
| 31 | DeepSeek V4 Pro 0813 DeepSeek · Reasoning, Max Effort | 36.0 | 36 | $0.150 / $3.50 |
| 32 | DeepSeek V4 Flash 0731 DeepSeek · Reasoning, Max Effort | 34.3 | 196 | $0.140 / $0.280 |
| 33 | Gemini 3.6 Flash Google · high | 34.0 | 13 | $1.35 / $6.75 |
| 34 | GLM-5.2 Z.ai · max | 33.7 | 30 | $0.180 / $4.00 |
| 35 | Muse Spark 1.1 Meta · xhigh | 33.7 | 17 | $1.25 / $4.25 |
| 36 | Qwen3.8 27B Qwen · xhigh | 33.7 | 30 | $0.025 / $4.35 |
| 37 | Gemini 3.5 Flash Google · high | 32.6 | 5.4 | $2.70 / $16.20 |
| 38 | DeepSeek V4 Pro 0424 DeepSeek · Reasoning, Max Effort | 30.4 | 26 | $0.400 / $3.50 |
| 39 | Claude Sonnet 4.6 Anthropic · Adaptive Reasoning, Max Effort | 30.1 | 5.0 | $3.00 / $15.00 |
| 40 | Gemini 3.1 Pro Preview | 29.7 | 3.7 | $3.60 / $21.60 |
| 41 | Qwen3.7 Max Qwen | 29.5 | 13 | $1.48 / $4.42 |
| 42 | MiniMax-M3 MiniMax | 29.2 | 28 | $0.600 / $2.40 |
| 43 | Kimi K2.6 MoonshotAI | 27.0 | 18 | $0.855 / $3.60 |
| 44 | GLM-5.1 Z.ai · Reasoning | 26.1 | 13 | $1.33 / $4.18 |
| 45 | MiMo-V2.5-Pro Xiaomi | 26.0 | 40 | $0.522 / $1.04 |
| 46 | Kimi K2.7 Code MoonshotAI | 25.8 | 15 | $0.950 / $4.00 |
| 47 | Hy3 Tencent | 25.3 | 89 | $0.180 / $0.600 |
| 48 | Qwen3.7 Plus Qwen | 25.2 | 45 | $0.320 / $1.28 |
| 49 | Inkling Thinking Machines · xhigh | 25.0 | Gratis | $0 / $0 |
| 50 | Grok 4.3 SpaceXAI · high | 24.9 | 20 | $1.00 / $2.00 |
Valor = puntuación ÷ precio combinado por 1M tokens. Cuanto mayor, mejor.
Puntuaciones de inteligencia, programación y agente lado a lado, con valor por 1M tokens.
Las puntuaciones provienen de Artificial Analysis, servidas a través de OpenRouter, en tres índices independientes: inteligencia para razonamiento y conocimiento general, programación para tareas reales de código, y agente para el uso de herramientas en varios pasos y la ejecución autónoma. Suelen discrepar a menudo — el modelo más inteligente no siempre es el mejor programando, y unas puntuaciones de agente fuertes suelen traer un precio acorde. Poner las tres junto a la columna de precios es lo que convierte «¿deberíamos cambiar?» en una decisión con evidencias detrás.
Cómo usar
- 1Cambia el benchmark (inteligencia / programación / agente) y la tabla se reclasifica por esa puntuación.
- 2Ordena por valor para ver cuántos puntos compra cada dólar por 1M tokens.
- 3Lee hacia la columna de precios para encontrar la mejor puntuación dentro de tu presupuesto.