Matriz de Seleccion de Modelos de AI

Overview

Esta matriz compara LLM models across cost, latency, context window, accuracy, y task suitability. Usala para selectar el right model para tu use case y budget. Updatea pricing y benchmarks a medida que new models se releasean.

---

1. Model Comparison Table

Model              | Input $/1K | Output $/1K | Context | Latency p50 | Multimodal
───────────────────┼────────────┼─────────────┼─────────┼─────────────┼───────────
GPT-4o             | $0.00250   | $0.01000    | 128K    | 800ms       | Yes
GPT-4o-mini        | $0.00015   | $0.00060    | 128K    | 300ms       | Yes
Claude 3.5 Sonnet  | $0.00300   | $0.01500    | 200K    | 1000ms      | Yes
Claude 3 Haiku     | $0.00025   | $0.00125    | 200K    | 400ms       | Yes
Llama 3.1 70B      | $0.00059   | $0.00079    | 128K    | 600ms       | No
Llama 3.1 8B       | $0.00005   | $0.00008    | 128K    | 200ms       | No
Gemini 1.5 Pro     | $0.00125   | $0.00500    | 2M      | 1200ms      | Yes
Gemini 1.5 Flash   | $0.000075  | $0.00030    | 1M      | 400ms       | Yes
Mistral Large      | $0.00200   | $0.00600    | 128K    | 700ms       | No
Mistral Small      | $0.00020   | $0.00060    | 128K    | 300ms       | No

---

2. Cost per 1000 Queries

Assumptions: 1500 input tokens, 500 output tokens per query

Model              | Cost per query | Cost per 1K queries | Monthly (10K q/day)
───────────────────┼────────────────┼─────────────────────┼────────────────────
GPT-4o             | $0.00875       | $8.75               | $2,625.00
GPT-4o-mini        | $0.00053       | $0.53               | $157.50
Claude 3.5 Sonnet  | $0.01125       | $11.25              | $3,375.00
Claude 3 Haiku     | $0.00094       | $0.94               | $281.25
Llama 3.1 70B      | $0.00129       | $1.29               | $386.25
Llama 3.1 8B       | $0.00012       | $0.12               | $35.40
Gemini 1.5 Pro     | $0.00438       | $4.38               | $1,312.50
Gemini 1.5 Flash   | $0.00021       | $0.21               | $63.75
Mistral Large      | $0.00600       | $6.00               | $1,800.00
Mistral Small      | $0.00060       | $0.60               | $180.00

---

3. Task-Based Recommendations

3.1 Classification

Task: Assignar input a uno de N categories
Requirements: Consistency, low cost, low latency

Recommended models (en order):
1. GPT-4o-mini     — Best cost/accuracy ratio, 128K context, fast
2. Claude 3 Haiku  — Good accuracy, slightly mas expensive
3. Gemini 1.5 Flash — Cheapest option, good para simple categories
4. Llama 3.1 8B    — Self-hosted option, lowest cost

Not recommended:
- GPT-4o          — Overkill para classification, 15x mas expensive
- Claude 3.5      — Overkill, 12x mas expensive que mini

3.2 Information Extraction

Task: Extraer structured data (entities, relationships, fields) de text
Requirements: Accuracy, structured output, schema adherence

Recommended models (en order):
1. GPT-4o          — Best structured output, JSON mode, high accuracy
2. Claude 3.5      — Strong en extraction, good con complex schemas
3. GPT-4o-mini     — Good para simple extraction, 15x cheaper
4. Gemini 1.5 Pro  — Good para long-document extraction (2M context)

Considerations:
- Usa JSON mode / structured output cuando available
- Valida output schema en application code
- Para long documents (>100K tokens), usa Gemini 1.5 Pro o Claude 3.5

3.3 Summarization

Task: Condensar long text en un shorter summary
Requirements: Accuracy, coherence, context window para long inputs

Recommended models (en order):
1. Claude 3.5 Sonnet — Best summarization quality, 200K context
2. GPT-4o            — Strong summarization, 128K context
3. Gemini 1.5 Pro    — Best para very long documents (2M context)
4. Claude 3 Haiku    — Good para short documents, lower cost

Considerations:
- Para documents < 10K tokens: any model trabaja well
- Para documents 10K-100K tokens: GPT-4o o Claude 3.5
- Para documents > 100K tokens: Gemini 1.5 Pro (2M context)

3.4 Code Generation y Review

Task: Generar, reviewar, o refactorear code
Requirements: Code accuracy, language support, reasoning ability

Recommended models (en order):
1. Claude 3.5 Sonnet — Best code reasoning, strong en refactoring
2. GPT-4o            — Excellent code generation, broad language support
3. GPT-4o-mini       — Good para simple code tasks, lower cost
4. Llama 3.1 70B     — Good open-source option para self-hosting

Considerations:
- Para complex refactoring: Claude 3.5 Sonnet
- Para boilerplate generation: GPT-4o-mini
- Para code review: GPT-4o (best en identifying subtle bugs)
- Para self-hosting: Llama 3.1 70B con code-specific fine-tunes

3.5 RAG (Retrieval-Augmented Generation)

Task: Answer questions basado en retrieved context
Requirements: Faithfulness, context utilization, citation accuracy

Recommended models (en order):
1. GPT-4o            — Best faithfulness, sigue context instructions well
2. Claude 3.5 Sonnet — Strong en synthesizing multiple chunks
3. Gemini 1.5 Pro    — Best para large context windows (many chunks)
4. GPT-4o-mini       — Good para simple Q&A, lower cost

Considerations:
- Faithfulness es critical — usa temperature 0
- Para multi-chunk synthesis: Claude 3.5 o GPT-4o
- Para large context (20+ chunks): Gemini 1.5 Pro
- Para cost-sensitive RAG: GPT-4o-mini con fewer chunks

3.6 Agent / Tool Use

Task: Multi-step reasoning con tool calls
Requirements: Tool selection accuracy, reasoning, instruction following

Recommended models (en order):
1. Claude 3.5 Sonnet — Best tool use, strong reasoning, function calling
2. GPT-4o            — Excellent tool use, parallel function calling
3. GPT-4o-mini       — Good para simple agents (1-3 tools)
4. Llama 3.1 70B     — Open-source option con function calling

Considerations:
- Para complex agents (5+ tools): Claude 3.5 o GPT-4o
- Para simple agents (1-3 tools): GPT-4o-mini
- Para self-hosted agents: Llama 3.1 70B
- Testea tool selection accuracy antes de deployear

---

4. Decision Framework

Step 1: Cuál es el task type?
  Classification → Empeza con GPT-4o-mini
  Extraction → Empeza con GPT-4o
  Summarization → Empeza con Claude 3.5
  Code → Empeza con Claude 3.5
  RAG → Empeza con GPT-4o
  Agent → Empeza con Claude 3.5

Step 2: Cuál es el input size?
  < 10K tokens → Any model trabaja
  10K-100K tokens → GPT-4o, Claude 3.5, o Llama 3.1
  100K-500K tokens → Claude 3.5 (200K) o Gemini 1.5 Pro (2M)
  > 500K tokens → Gemini 1.5 Pro (2M context)

Step 3: Cuál es tu budget per 1000 queries?
  < $1 → GPT-4o-mini, Gemini Flash, Llama 8B
  $1-5 → GPT-4o, Gemini Pro, Llama 70B
  $5-15 → Claude 3.5, GPT-4o, Mistral Large
  > $15 → Any model (cost no es el constraint)

Step 4: Necesitas self-hostear?
  Si → Llama 3.1 (8B o 70B dependiendo del hardware)
  No → Any cloud model

Step 5: Necesitas multimodal (image input)?
  Si → GPT-4o, Claude 3.5, Gemini 1.5
  No → Any model

---

5. Benchmark Results

Benchmark: MMLU (general knowledge)
Model              | Score
───────────────────┼──────
GPT-4o             | 88.7%
Claude 3.5 Sonnet  | 88.3%
Gemini 1.5 Pro     | 85.9%
Llama 3.1 70B      | 82.0%
GPT-4o-mini        | 81.0%
Claude 3 Haiku     | 75.2%
Gemini 1.5 Flash   | 77.9%
Llama 3.1 8B       | 68.5%

Benchmark: HumanEval (code generation)
Model              | Score
───────────────────┼──────
Claude 3.5 Sonnet  | 92.0%
GPT-4o             | 90.2%
Gemini 1.5 Pro     | 84.1%
Llama 3.1 70B      | 80.5%
GPT-4o-mini        | 87.0%
Claude 3 Haiku     | 75.1%

Benchmark: Tool Use (function calling accuracy)
Model              | Score
───────────────────┼──────
Claude 3.5 Sonnet  | 95.0%
GPT-4o             | 93.0%
GPT-4o-mini        | 85.0%
Llama 3.1 70B      | 78.0%

---

6. Fallback Strategy

# Multi-model fallback chain
FALLBACK_CHAIN = [
    {"model": "gpt-4o", "timeout": 30, "retry": 2},
    {"model": "claude-3-5-sonnet", "timeout": 30, "retry": 1},
    {"model": "gpt-4o-mini", "timeout": 15, "retry": 1},
]

async def call_with_fallback(prompt: str, **kwargs):
    for config in FALLBACK_CHAIN:
        try:
            response = await call_llm(
                model=config["model"],
                prompt=prompt,
                timeout=config["timeout"],
                retry_count=config["retry"],
                **kwargs,
            )
            return response
        except (TimeoutError, RateLimitError, ServiceUnavailableError):
            continue  # Try next model
    
    raise Exception("All models in fallback chain failed")

Puntos Clave

• Aplica matriz de seleccion de modelos de ai cuando necesites una solución práctica para tu caso de uso.
• Monitorea el rendimiento después de implementar; mide latencia, errores y uso de recursos antes y después.
• Revisa la sección de Troubleshooting ante errores comunes; la mayoría tienen causa raíz documentada con solución.
• Mantén dependencias actualizadas y ejecuta tests en CI para prevenir regresiones en producción.

See Also

• Complete Guide to LLM Prompt Engineering
• Complete Guide to LangChain in Production
• Complete Guide to LLM Application Architecture
• Complete Guide to LLM Cost Optimization
• Complete Guide to LLM Evaluation

Troubleshooting

• Model outputs are inconsistent: set temperature to 0 for deterministic tasks, use seed where supported, and version the prompt.
• Prompt injection leaks context: separate user input from system instructions.
• High token costs: cache embeddings, summarize long context, and choose smaller models for simple tasks.
• Retrieval returns irrelevant chunks: tune chunk size, overlap, and metadata filters. Evaluate retrieval metrics separately from generation.
• Evaluation scores do not match human judgment: Human review is still the ground truth.

Errores Comunes en Producción

• Dejar campos requeridos vacíos o usar respuestas vagas de una palabra.
• Llenar el documento una vez y nunca actualizarlo cuando cambia el alcance o las decisiones.
• Guardar el documento donde el equipo no lo busque durante incidentes o revisiones.
• No asignar un responsable, fecha límite o cadencia de revisión.
• Copiar texto base sin eliminar secciones que no aplican.
• Saltar el control de versiones, lo que impide rollback y responsabilidad.
• No vincular el documento con decisiones relacionadas o acciones de seguimiento.
• Evitar revisiones trimestrales que retirarían secciones obsoletas o sin uso.