📊 Benchmarks de Coding
Ranking de modelos de IA en los principales benchmarks de programación. Datos recopilados de fuentes públicas a agosto de 2026.
⚠️ Los benchmarks tienen limitaciones: pueden ser "harness-tuned", contaminados o no representar tu caso de uso específico. Úsalos como orientación, no como verdad absoluta. Verifica con evaluaciones independientes como LMSYS Chatbot Arena o Artificial Analysis.
🏆 Ranking SWE-bench Pro
El benchmark más relevante para evaluar capacidad de ingeniería de software real (resolución de issues en repos multi-archivo).
1
Claude Opus 4.8 Anthropic
79.4%
3
GPT-5.5 OpenAI
75.2%
4
Gemini 3.1 Pro Google
71.8%
5
Kimi K2.6 Moonshot
71.6%
6
DeepSeek V4 Pro DeepSeek
68.5%
Tabla completa de datos
| Modelo | Empresa | Contexto | $/M entrada | SWE-bench Pro | HumanEval | Open-weight |
|---|---|---|---|---|---|---|
| Claude Opus 4.8 | Anthropic | 200K | $5 | 79.4% | — | — |
| GLM-5 | Zhipu | 200K | $1 | 77.8% | — | ✓ |
| GPT-5.5 | OpenAI | 400K | $5 | 75.2% | 96.2% | — |
| Gemini 3.1 Pro | 2M | $2 | 71.8% | 94.6% | — | |
| Kimi K2.6 | Moonshot | 262K | $0.95 | 71.6% | — | — |
| DeepSeek V4 Pro | DeepSeek | 128K | $0.435 | 68.5% | — | ✓ |
| Claude Sonnet 4.6 | Anthropic | 1M | $3 | — | — | — |
| Claude Haiku 4.5 | Anthropic | 200K | $1 | — | — | — |
| GPT-5.5 Pro | OpenAI | 400K | $30 | — | — | — |
| GPT-5.4 | OpenAI | 400K | $2.5 | — | — | — |
| GPT-5.3-Codex | OpenAI | 400K | $1.75 | — | — | — |
| GPT-4.1 nano | OpenAI | 1M | $0.1 | — | 0.9% | — |
| Gemini 3.5 Flash | 1M | $1.5 | — | — | — | |
| Gemini 3 Flash | 1M | $0.5 | — | — | — | |
| Gemini 3.1 Flash-Lite | 1M | $0.25 | — | — | — | |
| Phi-4 (familia) | Microsoft | 16K | $0.125 | — | 0.9% | ✓ |
| MAI-Code-1-Flash | Microsoft | 128K | — | — | — | — |
| Grok 3 | xAI | 131K | $3 | — | 0.9% | — |
| Grok 3 Mini | xAI | 131K | $0.3 | — | — | — |
| Llama 4 Maverick | Meta | 1M | $0.27 | — | 0.9% | ✓ |
| Mistral Large 3 | Mistral | 256K | $0.5 | — | — | ✓ |
| Mistral Small | Mistral | 128K | $0.2 | — | 0.8% | ✓ |
| DeepSeek V4 Flash | DeepSeek | 128K | $0.14 | — | — | — |
| DeepSeek V3 | DeepSeek | 128K | $0.27 | — | 0.9% | ✓ |
| Qwen 3.6 Plus | Alibaba | 1M | $0.5 | — | — | — |
| Qwen3.5 (open-weight) | Alibaba | 131K | $0.3 | — | — | ✓ |
| MiniMax M2.5 | MiniMax | 1M | $0.3 | — | — | — |
| MiMo-V2-Pro | Xiaomi | 256K | $0 | — | — | ✓ |
| ERNIE 5 | Baidu | 128K | $0.4 | — | — | — |
| Hunyuan Large | Tencent | 256K | $0.4 | — | — | ✓ |
| Step 3.5 Flash | StepFun | 128K | $0.1 | — | — | — |
| GPT-5.6 Sol (xhigh) | OpenAI | 1M | $5 | — | — | — |
| Claude Opus 5 (max) | Anthropic | 1M | $5 | — | — | — |
| GPT-5.6 Sol (max) | OpenAI | 1M | $5 | — | — | — |
| GPT-5.6 Sol (high) | OpenAI | 1M | $5 | — | — | — |
| Claude Opus 5 (xhigh) | Anthropic | 1M | $5 | — | — | — |
| GPT-5.6 Terra (max) | OpenAI | 1M | $2 | — | — | — |
| Claude Opus 5 (high) | Anthropic | 1M | $5 | — | — | — |
| Claude Fable 5 (with fallback) | Anthropic | 1M | $10 | — | — | — |
| GPT-5.6 Sol (medium) | OpenAI | 1M | $5 | — | — | — |
| Kimi K3 (max) | Moonshot | 1.048576M | $3 | — | — | ✓ |
| Claude Opus 5 (medium) | Anthropic | 1M | $5 | — | — | — |
| Grok 4.5 (high) | SpaceXAI | 500K | $2 | — | — | — |
| Kimi K3 (low) | Moonshot | 1.048576M | $3 | — | — | ✓ |
| Claude Sonnet 5 (max) | Anthropic | 1M | $2 | — | — | — |
| GPT-5.6 Luna (max) | OpenAI | 1M | $0.2 | — | — | — |
| Muse Spark 1.1 (xhigh) | Meta | 1.048576M | $1.25 | — | — | — |
| GPT-5.6 Terra (xhigh) | OpenAI | 1M | $2 | — | — | — |
| GPT-5.6 Sol (low) | OpenAI | 1M | $5 | — | — | — |
| Gemini 3.6 Flash | 1M | $1.5 | — | — | — | |
| Gemini 3.1 Pro Preview | 1M | $2 | — | — | — | |
| GLM-5.2 (max) | Zhipu | 1M | $1.4 | — | — | ✓ |
| GPT-5.6 Luna (xhigh) | OpenAI | 1M | $0.2 | — | — | — |
| GPT-5.6 Terra (high) | OpenAI | 1M | $2 | — | — | — |
| Claude Opus 5 (low) | Anthropic | 1M | $5 | — | — | — |
| Claude Sonnet 5 (Non-reasoning) | Anthropic | 1M | $2 | — | — | — |
| Qwen3.7 Max | Alibaba | 1M | $2.5 | — | — | — |
| GPT-5.6 Sol (Non-reasoning) | OpenAI | 1M | $5 | — | — | — |
| GPT-5.6 Terra (medium) | OpenAI | 1M | $2 | — | — | — |
| GPT-5.6 Luna (high) | OpenAI | 1M | $0.2 | — | — | — |
| Motif 3 (Beta) | MotifTechnologies | 262K | $0 | — | — | — |
| Kimi K2.7 Code | Moonshot | 256K | $0.95 | — | — | ✓ |
| MiMo-V2.5-Pro | Xiaomi | 1M | $0.435 | — | — | ✓ |
| KAT-Coder-Pro V2 | KwaiKAT | 256K | $0.3 | — | — | — |
| Nex-N2-Pro | NexAGI | 262K | $0.5 | — | — | ✓ |
| Hy3 | Tencent | 256K | $0.136 | — | — | ✓ |
| Agnes 2.5 Pro Alpha | SapiensAI | 1M | $0.45 | — | — | — |
| DeepSeek V4 Pro (high) | DeepSeek | 1M | $0.435 | — | — | ✓ |
| Muse Spark | Meta | 262K | $0 | — | — | — |
| MiniMax-M3 | MiniMax | 1M | $0.3 | — | — | ✓ |
| GPT-5.6 Terra (low) | OpenAI | 1M | $2 | — | — | — |
| MiMo-V2.5 | Xiaomi | 1M | $0.14 | — | — | ✓ |
| Qwen3.7 Plus | Alibaba | 1M | $0.4 | — | — | — |
| Qwen3.6 Plus | Alibaba | 1M | $0.5 | — | — | — |
| Qwen3.6 27B | Alibaba | 262K | $0.6 | — | — | ✓ |
| Inkling Small | ThinkingMachines | 1M | $0.3 | — | — | ✓ |
| JT-4.1 Flash 236B A21B | ChinaMobile | 256K | $0 | — | — | — |
| GPT-5.6 Terra (Non-reasoning) | OpenAI | 1M | $2 | — | — | — |
Claude Opus 4.8
Anthropic
200K
$5
79.4%
—
—
GLM-5
Zhipu
200K
$1
77.8%
—
✓
GPT-5.5
OpenAI
400K
$5
75.2%
96.2%
—
Gemini 3.1 Pro
Google
2M
$2
71.8%
94.6%
—
Kimi K2.6
Moonshot
262K
$0.95
71.6%
—
—
DeepSeek V4 Pro
DeepSeek
128K
$0.435
68.5%
—
✓
Claude Sonnet 4.6
Anthropic
1M
$3
—
—
—
Claude Haiku 4.5
Anthropic
200K
$1
—
—
—
GPT-5.5 Pro
OpenAI
400K
$30
—
—
—
GPT-5.4
OpenAI
400K
$2.5
—
—
—
GPT-5.3-Codex
OpenAI
400K
$1.75
—
—
—
GPT-4.1 nano
OpenAI
1M
$0.1
—
0.9%
—
Gemini 3.5 Flash
Google
1M
$1.5
—
—
—
Gemini 3 Flash
Google
1M
$0.5
—
—
—
Gemini 3.1 Flash-Lite
Google
1M
$0.25
—
—
—
Phi-4 (familia)
Microsoft
16K
$0.125
—
0.9%
✓
MAI-Code-1-Flash
Microsoft
128K
—
—
—
—
Grok 3
xAI
131K
$3
—
0.9%
—
Grok 3 Mini
xAI
131K
$0.3
—
—
—
Llama 4 Maverick
Meta
1M
$0.27
—
0.9%
✓
Mistral Large 3
Mistral
256K
$0.5
—
—
✓
Mistral Small
Mistral
128K
$0.2
—
0.8%
✓
DeepSeek V4 Flash
DeepSeek
128K
$0.14
—
—
—
DeepSeek V3
DeepSeek
128K
$0.27
—
0.9%
✓
Qwen 3.6 Plus
Alibaba
1M
$0.5
—
—
—
Qwen3.5 (open-weight)
Alibaba
131K
$0.3
—
—
✓
MiniMax M2.5
MiniMax
1M
$0.3
—
—
—
MiMo-V2-Pro
Xiaomi
256K
$0
—
—
✓
ERNIE 5
Baidu
128K
$0.4
—
—
—
Hunyuan Large
Tencent
256K
$0.4
—
—
✓
Step 3.5 Flash
StepFun
128K
$0.1
—
—
—
GPT-5.6 Sol (xhigh)
OpenAI
1M
$5
—
—
—
Claude Opus 5 (max)
Anthropic
1M
$5
—
—
—
GPT-5.6 Sol (max)
OpenAI
1M
$5
—
—
—
GPT-5.6 Sol (high)
OpenAI
1M
$5
—
—
—
Claude Opus 5 (xhigh)
Anthropic
1M
$5
—
—
—
GPT-5.6 Terra (max)
OpenAI
1M
$2
—
—
—
Claude Opus 5 (high)
Anthropic
1M
$5
—
—
—
Claude Fable 5 (with fallback)
Anthropic
1M
$10
—
—
—
GPT-5.6 Sol (medium)
OpenAI
1M
$5
—
—
—
Kimi K3 (max)
Moonshot
1.048576M
$3
—
—
✓
Claude Opus 5 (medium)
Anthropic
1M
$5
—
—
—
Grok 4.5 (high)
SpaceXAI
500K
$2
—
—
—
Kimi K3 (low)
Moonshot
1.048576M
$3
—
—
✓
Claude Sonnet 5 (max)
Anthropic
1M
$2
—
—
—
GPT-5.6 Luna (max)
OpenAI
1M
$0.2
—
—
—
Muse Spark 1.1 (xhigh)
Meta
1.048576M
$1.25
—
—
—
GPT-5.6 Terra (xhigh)
OpenAI
1M
$2
—
—
—
GPT-5.6 Sol (low)
OpenAI
1M
$5
—
—
—
Gemini 3.6 Flash
Google
1M
$1.5
—
—
—
Gemini 3.1 Pro Preview
Google
1M
$2
—
—
—
GLM-5.2 (max)
Zhipu
1M
$1.4
—
—
✓
GPT-5.6 Luna (xhigh)
OpenAI
1M
$0.2
—
—
—
GPT-5.6 Terra (high)
OpenAI
1M
$2
—
—
—
Claude Opus 5 (low)
Anthropic
1M
$5
—
—
—
Claude Sonnet 5 (Non-reasoning)
Anthropic
1M
$2
—
—
—
Qwen3.7 Max
Alibaba
1M
$2.5
—
—
—
GPT-5.6 Sol (Non-reasoning)
OpenAI
1M
$5
—
—
—
GPT-5.6 Terra (medium)
OpenAI
1M
$2
—
—
—
GPT-5.6 Luna (high)
OpenAI
1M
$0.2
—
—
—
Motif 3 (Beta)
MotifTechnologies
262K
$0
—
—
—
Kimi K2.7 Code
Moonshot
256K
$0.95
—
—
✓
MiMo-V2.5-Pro
Xiaomi
1M
$0.435
—
—
✓
KAT-Coder-Pro V2
KwaiKAT
256K
$0.3
—
—
—
Nex-N2-Pro
NexAGI
262K
$0.5
—
—
✓
Hy3
Tencent
256K
$0.136
—
—
✓
Agnes 2.5 Pro Alpha
SapiensAI
1M
$0.45
—
—
—
DeepSeek V4 Pro (high)
DeepSeek
1M
$0.435
—
—
✓
Muse Spark
Meta
262K
$0
—
—
—
MiniMax-M3
MiniMax
1M
$0.3
—
—
✓
GPT-5.6 Terra (low)
OpenAI
1M
$2
—
—
—
MiMo-V2.5
Xiaomi
1M
$0.14
—
—
✓
Qwen3.7 Plus
Alibaba
1M
$0.4
—
—
—
Qwen3.6 Plus
Alibaba
1M
$0.5
—
—
—
Qwen3.6 27B
Alibaba
262K
$0.6
—
—
✓
Inkling Small
ThinkingMachines
1M
$0.3
—
—
✓
JT-4.1 Flash 236B A21B
ChinaMobile
256K
$0
—
—
—
GPT-5.6 Terra (Non-reasoning)
OpenAI
1M
$2
—
—
—
¿Qué mide cada benchmark?
SWE-bench Pro
Benchmark de referencia para ingeniería de software: el modelo debe resolver issues reales de GitHub en repos multi-archivo. Mide capacidad agéntica completa: leer código, planificar, implementar y verificar.
Ver fuente →HumanEval
Benchmark clásico de generación de código Python: 164 problemas de programación con tests unitarios. Mide la capacidad de escribir funciones correctas a partir de docstrings.
Ver fuente →LiveCodeBench
Benchmark dinámico con problemas de competiciones de programación recientes (post-entrenamiento). Evita contaminación de datos.
Ver fuente →💡 Cómo interpretar los benchmarks
Match modelo → tarea
Un modelo top en SWE-bench puede no ser el mejor para tu caso. Si haces mucho prototyping rápido, un modelo "flash" barato puede ser más productivo que el #1 del ranking.
Cuidado con el benchmark gaming
Los vendors optimizan para benchmarks específicos. SWE-bench Pro y LiveCodeBench son más difíciles de gamear, pero no son inmunes. Valida siempre con tu propio código.
Actualiza trimestralmente
Los modelos se actualizan cada 2-3 meses. Un ranking de enero puede ser obsoleto en abril. Fuentes fiables: Artificial Analysis, LiveBench, SWE-bench.