AI Models & Routing

Understand how Apex AI Brain selects and routes tasks between different AI models.

Supported Models

Claude (Anthropic) - Default Primary

Best for: Complex architecture, creative solutions

Models:
├─ Claude Opus        (200k context) - Most capable
  ├─ Claude Sonnet      (200k context) - Balanced  
  └─ Claude Haiku       (200k context) - Fast & cheap

Pricing:
├─ Opus:   $15 / 1M input,  $75 / 1M output
├─ Sonnet: $3  / 1M input,  $15 / 1M output
└─ Haiku:  $0.80 / 1M input, $4 / 1M output

Strengths:

Excellent code generation
Strong reasoning
Good context understanding
Architectural design

Google Gemini - Fast & Efficient

Best for: Quick analysis, cost-effective solutions

Models:
├─ Gemini Pro       (30k context) - General purpose
└─ Gemini Pro Vision (30k context) - With image input

Pricing:
├─ Pro:      $0.50 / 1M input,  $1.50 / 1M output
└─ Pro Vision: $1 / 1M input,    $3 / 1M output

Strengths:

Very fast responses
Cost-effective
Good for code review
Parallel processing friendly

GPT-4 / GPT-4 Turbo - Experimental

Best for: Advanced reasoning, specialized tasks

Models:
├─ GPT-4 Turbo   (128k context) - Latest & fastest
└─ GPT-4         (8k context)   - Legacy

Pricing:
├─ Turbo: $10 / 1M input, $30 / 1M output
└─ GPT-4: $30 / 1M input, $60 / 1M output

Model Selection Strategy

Automatic Routing

Apex AI Brain automatically selects the best model based on:

┌────────────────────┐
│  Task Analysis     │
├────────────────────┤
│ Complexity Level   │
│ Output Type        │
│ Context Size       │
│ Budget Constraint  │
└────────────┬───────┘
             ↓
┌────────────────────┐
│  Model Scoring     │
├────────────────────┤
│ Score each model   │
│ Weight by factors  │
│ Select best fit    │
└────────────┬───────┘
             ↓
┌────────────────────┐
│  Execute &Track    │
└────────────────────┘

Model Assignment by Agent

Agent	Default	Fallback	Reason
apex-architect	Opus	Sonnet	Complex architecture
lwc-builder	Sonnet	Gemini	Balanced performance
apex-reviewer	Opus	Sonnet	Deep analysis needed
soql-optimizer	Gemini	Haiku	Quick optimization
schema-analyst	Sonnet	Gemini	Structured design
security-agent	Opus	Sonnet	Thorough review
test-writer	Sonnet	Gemini	Good patterns
deployment-agent	Gemini	Sonnet	Fast planning

Manual Model Selection

Set Default Model

# Global default
sfai /model set claude

# For specific command
sfai /spawn --model gemini soql-optimizer "Optimize queries"

Model-Specific Options

# Use specific Claude model
sfai /spawn --model opus apex-architect "..."

# Use Gemini for speed
sfai /spawn --model gemini lwc-builder "..."

# Use GPT-4 for advanced reasoning
sfai /spawn --model gpt4 debug-agent "..."

Context Window Management

Context Sizes

Model              Context Size    Cost Factor
─────────────────────────────────────────────
Claude Opus        200k tokens     3.0x
Claude Sonnet      200k tokens     1.0x (baseline)
Claude Haiku       200k tokens     0.25x
Gemini Pro         30k tokens      0.25x
GPT-4 Turbo        128k tokens     2.0x

Context Optimization

When context is too large:

Summarize code
Extract key patterns
Remove comments
Strip boilerplate
Keep core logic

Example:

Original: 50k tokens (too large)
    ↓
Summarized: 5k tokens (sends core concepts)
    ↓
Response size reduced by 90%
    ↓
Cost reduced by 90%

Token Tracking & Cost Management

Token Usage Report

sfai /tokens --breakdown

Sample output:

Token Usage This Session
========================

By Model:
  Claude Opus:    30,000  →  $0.45
  Claude Sonnet:  50,000  →  $0.15
  Gemini Pro:     15,000  →  $0.08
  ───────────────────────────────
  Total:          95,000  →  $0.68

By Agent:
  apex-architect:   40,000  →  $0.30
  lwc-builder:      30,000  →  $0.18
  test-writer:      25,000  →  $0.20

Projected Monthly:
  Current:  $20.40
  Budget:   $100.00
  Status:    On track

Cost Optimization

Strategy 1: Use Cheaper Models for Simple Tasks

❌ WRONG:
apex-architect using Opus to optimize query
  Cost: $0.15 for quick task

 RIGHT:
soql-optimizer using Gemini to optimize query
  Cost: $0.02 for same task (7.5x cheaper!)

Strategy 2: Combine Fast + Complex

# 1. Fast agents first (Gemini)
sfai /spawn soql-optimizer --model gemini "Analyze queries"

# 2. Complex agents second (Claude Opus)
sfai /spawn apex-architect --model opus "Design architecture"

# 3. Final review (Claude Sonnet)
sfai /spawn apex-reviewer --model sonnet "Review code"

Total: More results, lower cost!

Strategy 3: Prompt Engineering

Shorter, focused prompts:

Less context sent
Faster processing
Lower token usage
Better responses

Model Comparison

Code Generation

Claude Opus:     95% quality, follows patterns perfectly
Claude Sonnet:   90% quality, good patterns
Gemini Pro:      85% quality, basic patterns
GPT-4 Turbo:     92% quality, creative solutions

Speed

Gemini Pro:      ⚡⚡⚡ Fastest (2-5s)
Claude Haiku:    ⚡⚡ Fast (5-10s)
Claude Sonnet:   ⚡ Medium (8-15s)
Claude Opus:     🐌 Slower (10-20s)
GPT-4 Turbo:     🐌 Slowest (15-30s)

Cost Efficiency

Claude Haiku:    💰 Cheapest (0.25x baseline)
Gemini Pro:      💰 Cheap (0.25x baseline)
Claude Sonnet:   💲 Baseline (1.0x)
Claude Opus:     💲💲 Expensive (3.0x)
GPT-4 Turbo:     💲💲 Most expensive (2.0x)

Advanced Model Configuration

Create Model Profile

In sfai.config.json:

{
  "modelProfiles": {
    "fast": {
      "primary": "gemini",
      "timeout": 5000,
      "retries": 1
    },
    "quality": {
      "primary": "claude-opus",
      "timeout": 20000,
      "retries": 3
    },
    "balanced": {
      "primary": "claude-sonnet",
      "fallback": "gemini",
      "timeout": 10000,
      "retries": 2
    }
  }
}

Use Profile

sfai /spawn --profile fast soql-optimizer "..."
sfai /spawn --profile quality apex-architect "..."

Model Fallback Strategy

Automatic Fallback Chain

Primary model fails
    ↓
Try fallback model 1
    ↓
Try fallback model 2
    ↓
Return best attempt or error

Configure Fallback

{
  "modelFallback": {
    "claude-opus": ["claude-sonnet", "gpt4"],
    "claude-sonnet": ["claude-haiku", "gemini"],
    "gemini": ["claude-haiku"],
    "gpt4": ["claude-opus"]
  }
}

Model-Specific Parameters

Claude-Specific

sfai /spawn --model claude \
  --temperature 0.7 \          # Creativity (0-1)
  --max-tokens 2000 \          # Output limit
  apex-architect "..."

Gemini-Specific

sfai /spawn --model gemini \
  --temperature 0.5 \
  --top-k 40 \                 # Diversity
  --top-p 0.95 \
  lwc-builder "..."

Cost Control

Monthly Budget Configuration

{
  "costControl": {
    "monthlyBudget": 100,
    "alertThreshold": 0.80,
    "hardLimit": true
  }
}

When budget exceeded:

alertThreshold: 0.80 → Warning at $80
hardLimit: true → Stop after $100

Per-Agent Budget

{
  "agentBudget": {
    "apex-architect": 30,
    "lwc-builder": 20,
    "test-writer": 15
  }
}

Performance & Cost Matrix

Use this to choose the best model:

Task	Recommended	Reason
Simple code review	Gemini	Fast & cheap
SOQL optimization	Gemini	Straightforward analysis
Architecture design	Claude Opus	Complex reasoning needed
Component building	Claude Sonnet	Balance quality & cost
Test generation	Claude Sonnet	Good test patterns
Security audit	Claude Opus	Thorough analysis
Debugging	Claude Sonnet	Good at finding issues
Deployment planning	Gemini	Quick planning

Monitoring

Real-time Monitor

watch -n 5 'sfai /tokens --live'

Daily Report

sfai /tokens --daily --export report.json

Analysis

# Which agents cost the most?
sfai /cost --by-agent

# Trend analysis
sfai /cost --trend 30days

# Projection
sfai /cost --projection end-of-month

Supported Models​

Claude (Anthropic) - Default Primary​

Google Gemini - Fast & Efficient​

GPT-4 / GPT-4 Turbo - Experimental​

Model Selection Strategy​

Automatic Routing​

Model Assignment by Agent​

Manual Model Selection​

Set Default Model​

Model-Specific Options​

Context Window Management​

Context Sizes​

Context Optimization​

Token Tracking & Cost Management​

Token Usage Report​

Cost Optimization​

Strategy 1: Use Cheaper Models for Simple Tasks​

Strategy 2: Combine Fast + Complex​

Strategy 3: Prompt Engineering​

Model Comparison​

Code Generation​

Speed​

Cost Efficiency​

Advanced Model Configuration​

Create Model Profile​

Use Profile​

Model Fallback Strategy​

Automatic Fallback Chain​

Configure Fallback​

Model-Specific Parameters​

Claude-Specific​

Gemini-Specific​

Cost Control​

Monthly Budget Configuration​

Per-Agent Budget​

Performance & Cost Matrix​

Monitoring​

Real-time Monitor​

Daily Report​

Analysis​