Model Hierarchy

Model Hierarchy

Route tasks to the cheapest model that can handle them. Most agent work is routine.

Core Principle

80% of agent tasks are janitorial. File reads, status checks, formatting, simple Q&A. These don't need expensive models. Reserve premium models for problems that actually require deep reasoning.

Model Tiers

Tier 1: Cheap ($0.10-0.50/M tokens)

Tier 2: Mid ($1-5/M tokens)

Tier 3: Premium ($10-75/M tokens)

Prices as of Feb 2026. Check provider docs for current rates.

Task Classification

Before executing any task, classify it:

ROUTINE → Use Tier 1

Characteristics:

• Single-step operations
• Clear, unambiguous instructions
• No judgment required
• Deterministic output expected

Examples:

• File read/write operations
• Status checks and health monitoring
• Simple lookups (time, weather, definitions)
• Formatting and restructuring text
• List operations (filter, sort, transform)
• API calls with known parameters
• Heartbeat and cron tasks
• URL fetching and basic parsing

MODERATE → Use Tier 2

Characteristics:

• Multi-step but well-defined
• Some synthesis required
• Standard patterns apply
• Quality matters but isn't critical

Examples:

• Code generation (standard patterns)
• Summarization and synthesis
• Draft writing (emails, docs, messages)
• Data analysis and transformation
• Multi-file operations
• Tool orchestration
• Code review (non-security)
• Search and research tasks

COMPLEX → Use Tier 3

Characteristics:

• Novel problem solving required
• Multiple valid approaches
• Nuanced judgment calls
• High stakes or irreversible
• Previous attempts failed

Examples:

• Multi-step debugging
• Architecture and design decisions
• Security-sensitive code review
• Tasks where cheaper model already failed
• Ambiguous requirements needing interpretation
• Long-context reasoning (>50K tokens)
• Creative work requiring originality
• Adversarial or edge-case handling

Decision Algorithm

function selectModel(task):
    # Rule 1: Escalation override
    if task.previousAttemptFailed:
        return nextTierUp(task.previousModel)
    
    # Rule 2: Explicit complexity signals
    if task.hasSignal("debug", "architect", "design", "security"):
        return TIER_3
    
    if task.hasSignal("write", "code", "summarize", "analyze"):
        return TIER_2
    
    # Rule 3: Default classification
    complexity = classifyTask(task)
    
    if complexity == ROUTINE:
        return TIER_1
    elif complexity == MODERATE:
        return TIER_2
    else:
        return TIER_3

Behavioral Rules

For Main Session

1. Default to Tier 2 for interactive work 2. Suggest downgrade when doing routine work: "This is routine - I can handle this on a cheaper model or spawn a sub-agent." 3. Request upgrade when stuck: "This needs more reasoning power. Switching to [premium model]."

For Sub-Agents

1. Default to Tier 1 unless task is clearly moderate+ 2. Batch similar tasks to amortize overhead 3. Report failures back to parent for escalation

For Automated Tasks

1. Heartbeats/monitoring → Always Tier 1 2. Scheduled reports → Tier 1 or 2 based on complexity 3. Alert responses → Start Tier 2, escalate if needed

Communication Patterns

When suggesting model changes, use clear language:

Downgrade suggestion:

"This looks like routine file work. Want me to spawn a sub-agent on DeepSeek for this? Same result, fraction of the cost."

Upgrade request:

"I'm hitting the limits of what I can figure out here. This needs Opus-level reasoning. Switching up."

Explaining hierarchy:

"I'm running the heavy analysis on Sonnet while sub-agents fetch the data on DeepSeek. Keeps costs down without sacrificing quality where it matters."

Cost Impact

Assuming 100K tokens/day average usage:

The 80/15/5 split:

• 80% routine tasks on Tier 1 (~$6)
• 15% moderate tasks on Tier 2 (~$7)
• 5% complex tasks on Tier 3 (~$6)

Result: 10x cost reduction vs pure premium, with equivalent quality on complex tasks.

Integration Examples

OpenClaw

# config.yml - set default model
model: anthropic/claude-sonnet-4

# In session, switch models
/model opus  # upgrade for complex task
/model deepseek  # downgrade for routine

# Spawn sub-agent on cheap model
sessions_spawn:
  task: "Fetch and parse these 50 URLs"
  model: deepseek

Claude Code

# In CLAUDE.md or project instructions
When spawning background agents, use claude-3-haiku for:
- File operations
- Simple searches  
- Status checks

Reserve claude-sonnet-4 for:
- Code generation
- Analysis tasks

General Agent Systems

def get_model_for_task(task_description: str) -> str:
    routine_signals = ['read', 'fetch', 'check', 'list', 'format', 'status']
    complex_signals = ['debug', 'architect', 'design', 'security', 'why']
    
    desc_lower = task_description.lower()
    
    if any(signal in desc_lower for signal in complex_signals):
        return "claude-opus-4"
    elif any(signal in desc_lower for signal in routine_signals):
        return "deepseek-v3"
    else:
        return "claude-sonnet-4"

Anti-Patterns

DON'T:

• Run heartbeats on Opus
• Use premium models for file I/O
• Keep expensive model when task is clearly routine
• Spawn sub-agents on premium models by default

DO:

• Start mid-tier, adjust based on task
• Spawn helpers on cheapest viable model
• Escalate explicitly when stuck
• Track cost per task type to optimize further

Extending This Skill

To customize for your use case:

1. Adjust tier definitions based on your provider/budget 2. Add domain-specific signals to classification rules 3. Track actual complexity vs predicted to improve heuristics 4. Set budget alerts to catch runaway premium usage