MODELOS IA / COMPARATIVA

Modelos IA

Benchmarks de los modelos que dan capacidades a los entornos. Datos de Artificial Analysis, filtrados por herramientas y tareas reales.

Benchmarks verificablesCATÁLOGO / 20 MODELOS

MODELOS / 001OPEN WEIGHTS · BENCHMARKS · CONTEXTO

COMPARATIVA AGÉNTICA

Los modelos más potentes para capacidades agénticas

20 modelos ordenados por el Agentic Index de Artificial Analysis, el composite que mide trabajo agéntico real (GDPval-AA v2 + 𝜏³-Banking). Incluye costo por tarea y tokens por segundo.

ACTUALIZADO25 ago 2026, 12:00 p.m.FUENTE: Artificial Analysis Agentic Index

#MODELOMÉTRICASCOSTO / VEL.

🥇

Claude Opus 5 (max)Anthropic

59.2

89.1%

INT

TAREA $5 / $25TOK 54/s54 t/s

🥈

GLM-5.3 (max)Z AI

59.1

83.9%

INT

TAREA $1.32TOK ——

🥉

Grok 4.6 (high)SpaceXAI

58.7

88.4%

INT

TAREA $1.95TOK 62/s54 t/s

GPT-5.6 Sol (max)OpenAI

57.8

88.0%

INT

TAREA $2.06TOK 74/s62 t/s

Qwen3.8 2.4T A95BAlibaba

57.1

82.0%

INT

TAREA $2 / $6TOK 27/s27 t/s

Claude Fable 5 (with fallback)Anthropic

56.6

84.6%

INT

TAREA $10 / $50TOK 66/s66 t/s

Kimi K3 (max)Kimi

54.3

85.0%

INT

TAREA $1.64TOK 35/s35 t/s

Qwen3.8 27B (xhigh)Alibaba

50.9

79.8%

INT

TAREA $0.38TOK 52/s52 t/s

GPT-5.6 Terra (max)OpenAI

50.2

88.0%

INT

TAREA $0.89TOK 122/s114 t/s

DeepSeek V4 Pro 0813 (max)DeepSeek

49.6

78.7%

INT

TAREA $0.37TOK 69/s69 t/s

Muse Spark 1.2 (xhigh)Meta

49.3

80.1%

INT

TAREA $0.61TOK ——

GPT-5.6 Luna (max)OpenAI

46.9

80.9%

INT

TAREA $0.08TOK 141/s139 t/s

Gemini 3.7 Flash (high)Google

45.1

85.8%

INT

TAREA $0.82TOK 358/s358 t/s

Motif 3Motif Technologies

37.6

74.9%

INT

TAREA $0 / $0TOK ——

MiniMax-M3MiniMax

36.1

65.2%

INT

TAREA $0.21TOK 126/s139 t/s

InklingThinking Machines

34.1

55.1%

INT

TAREA $0.44TOK 49/s47 t/s

Solar Open2 250BUpstage

27.8

44.2%

INT

TAREA $0 / $0TOK ——

Nemotron 3 UltraNVIDIA

27.5

53.9%

INT

TAREA $0.49TOK 168/s155 t/s

Gemini 3.5 Flash-LiteGoogle

27.2

53.6%

INT

TAREA $0.13TOK 382/s364 t/s

A.X-K2SK Telecom

25.7

39.0%

INT

TAREA $0 / $0TOK ——

El Agentic Index mide cuán bien un modelo trabaja con herramientas en tareas reales: tool use, planificación y ejecución multi-paso. TerminalBench v2.1 (TB) y el Intelligence Index (INT) acompañan como contexto; un modelo con alto INT pero bajo AG puede ser excelente conversando pero ineficaz operando agentes.

RANKING POR RUBRO

Los mejores en cada categoría

Top modelos por métrica individual. Cada rubro mide algo distinto — ningún modelo domina todos.

TerminalBench v2.1

1
GPT-5.6 Sol (xhigh)OpenAI
90%
2
Claude Opus 5 (Adaptive Reasoning, Max Effort)Anthropic
89%
3
Grok 4.6 (high)SpaceXAI
88%
4
GPT-5.6 Terra (max)OpenAI
88%
5
GPT-5.6 Sol (max)OpenAI
88%
6
Claude Opus 5 (Adaptive Reasoning, Xhigh Effort)Anthropic
88%
7
Grok 4.6 (xhigh)SpaceXAI
88%
8
Claude Opus 5 (Adaptive Reasoning, High Effort)Anthropic
88%
9
GPT-5.6 Sol (high)OpenAI
87%
10
GPT-5.6 Sol (medium)OpenAI
86%

HLE

Humanity's Last Exam

1
Claude Fable 5 (Adaptive Reasoning, Max Effort, Opus 4.8 Fallback)Anthropic
56%
2
Claude Opus 5 (Adaptive Reasoning, Max Effort)Anthropic
55%
3
Claude Opus 5 (Adaptive Reasoning, Xhigh Effort)Anthropic
54%
4
Claude Opus 5 (Adaptive Reasoning, High Effort)Anthropic
53%
5
Claude Opus 5 (Adaptive Reasoning, Medium Effort)Anthropic
51%
6
GPT-5.6 Sol (max)OpenAI
50%
7
Claude Opus 4.8 (Adaptive Reasoning, Max Effort)Anthropic
49%
8
Gemini 3.7 Flash (high)Google
48%
9
GPT-5.6 Sol (xhigh)OpenAI
47%
10
Gemini 3.1 Pro PreviewGoogle
47%

INT

Intelligence Index

1
Claude Opus 5 (Adaptive Reasoning, Max Effort)Anthropic
63
2
Claude Opus 5 (Adaptive Reasoning, Xhigh Effort)Anthropic
63
3
Claude Fable 5 (Adaptive Reasoning, Max Effort, Opus 4.8 Fallback)Anthropic
62
4
Claude Opus 5 (Adaptive Reasoning, High Effort)Anthropic
62
5
GPT-5.6 Sol (max)OpenAI
61
6
Grok 4.6 (high)SpaceXAI
61
7
Grok 4.6 (xhigh)SpaceXAI
60
8
Kimi K3 (max)Kimi OW
60
9
GLM-5.3 (max)Z AI OW
60
10
GPT-5.6 Sol (xhigh)OpenAI
59

COD

Coding Index

1
GPT-5.6 Sol (xhigh)OpenAI
78
2
Claude Opus 5 (Adaptive Reasoning, Max Effort)Anthropic
78
3
GPT-5.6 Sol (max)OpenAI
77
4
GPT-5.6 Sol (high)OpenAI
77
5
Claude Opus 5 (Adaptive Reasoning, Xhigh Effort)Anthropic
77
6
Grok 4.6 (high)SpaceXAI
77
7
GPT-5.6 Terra (max)OpenAI
77
8
Claude Fable 5 (Adaptive Reasoning, Max Effort, Opus 4.8 Fallback)Anthropic
77
9
Claude Opus 5 (Adaptive Reasoning, High Effort)Anthropic
77
10
GPT-5.6 Sol (medium)OpenAI
76

τ²

Tau2

1
GLM-5.2 (max)Z AI OW
99%
2
GLM-4.7-Flash (Reasoning)Z AI OW
99%
3
Claude Fable 5 (Adaptive Reasoning, Max Effort, Opus 4.8 Fallback)Anthropic
99%
4
GLM 5V Turbo (Reasoning)Z AI OW
99%
5
GLM-5-TurboZ AI OW
99%
6
GLM-5 (Reasoning)Z AI OW
98%
7
Grok 4.3 (high)SpaceXAI
98%
8
GLM-5.1 (Reasoning)Z AI OW
98%
9
Qwen3.6 PlusAlibaba
98%
10
GLM-5 (Non-reasoning)Z AI OW
97%

GPQA

1
Grok 4.6 (high)SpaceXAI
95%
2
Gemini 3.7 Flash (high)Google
95%
3
GPT-5.6 Sol (max)OpenAI
94%
4
Gemini 3.1 Pro PreviewGoogle
94%
5
Claude Opus 5 (Adaptive Reasoning, Xhigh Effort)Anthropic
94%
6
Claude Opus 5 (Adaptive Reasoning, High Effort)Anthropic
94%
7
Grok 4.6 (xhigh)SpaceXAI
94%
8
Grok 4.6 (medium)SpaceXAI
94%
9
Kimi K3 (max)Kimi OW
94%
10
Qwen3.8 2.4T A95BAlibaba OW
94%

OPEN WEIGHTS

Modelos con Parámetros Abiertos OpenWeight

Pesos descargables con licencia abierta. Para quienes quieren correr modelos en su propia infraestructura o usarlos con harnesses open source.

#MODELOMÉTRICASPRECIO / VEL.

🥇

Kimi K3 (max)OpenWeightKimi

85.0%

INT

COD

IN $3OUT $1535 t/s

🥈

GLM-5.3 (max)OpenWeightZ AI

83.9%

INT

COD

IN $1.4OUT $4.4—

🥉

Kimi K3 (low)OpenWeightKimi

82.4%

INT

COD

IN $3OUT $1532 t/s

Qwen3.8 2.4T A95BOpenWeightAlibaba

82.0%

INT

COD

IN $2OUT $627 t/s

Muse Spark 1.2 (xhigh)OpenWeightMeta

80.1%

INT

COD

IN $1.25OUT $4.25—

Qwen3.8 27B (xhigh)OpenWeightAlibaba

79.8%

INT

COD

IN $0.5OUT $352 t/s

DeepSeek V4 Pro 0813 (Reasoning, Max Effort)OpenWeightDeepSeek

78.7%

INT

COD

IN $1.32OUT $3.9669 t/s

DeepSeek V4 Flash 0731 (Reasoning, Max Effort)OpenWeightDeepSeek

78.7%

INT

COD

IN $0.44OUT $1.32109 t/s

GLM-5.2 (max)OpenWeightZ AI

77.9%

INT

COD

IN $1.4OUT $4.4—

Muse Spark 1.1 (xhigh)OpenWeightMeta

77.9%

INT

COD

IN $1.25OUT $4.25—

Kimi K2.7 CodeOpenWeightKimi

67.4%

INT

COD

IN $0.95OUT $443 t/s

Qwen3.8 27B (low)OpenWeightAlibaba

67.4%

INT

COD

IN $0.5OUT $358 t/s

GLOSARIO DE CAMPO

Ocho términos para leer la tabla sin perderte

Agentic Index: Score compuesto de Artificial Analysis que mide trabajo agéntico real: tool use, planificación, autonomía y ejecución multi-paso. Combina GDPval-AA v2 y 𝜏³-Banking. Escala 0–100.
Intelligence Index: Score compuesto de Artificial Analysis que mide capacidad general. Escala 0–100.
Coding Index: Score compuesto de capacidad de código. Escala 0–100.
TerminalBench v2.1: Mide si el modelo completa tareas reales en terminal: instalar dependencias, editar código, correr tests. La métrica agéntica más actualizada (sucesora de TerminalBench Hard).
Humanity's Last Exam (HLE): El examen de frontera de Artificial Analysis: preguntas durísimas que abarcan todo el conocimiento humano. La vara más exigente de capacidad general. Escala 0–1.
Tau2: Tareas multi-turno con herramientas. Mide retención, uso de tools y razonamiento en cadena.
GPQA: Preguntas de nivel postgrado en ciencia. Evalúa razonamiento profundo. Escala 0–1.
LiveCodeBench: Problemas de código reales y actualizados. Escala 0–1.
Open Weights: Modelos con pesos descargables. Puedes desplegarlos en tu propia infraestructura.
Velocidad: Tokens de salida por segundo. Más alto = respuestas más rápidas.

Ver señales relacionadas