LLM Model Selection

LLM Model Selection Skill

Choosing the right model for the task — power vs. cost vs. speed.

⚠️ Staleness Warning

This skill depends on rapidly evolving technology. Model capabilities, pricing, and availability change frequently.

Refresh triggers:

• New model announcements (Claude, GPT, Gemini, etc.)
• Significant pricing changes
• Context window expansions
• New capability tiers

Last validated: March 2026 (Claude 4.6 generation)

Check current state: Anthropic Models, OpenAI Models

The Core Question

Is Claude Opus 4.6 overkill?

Sometimes yes, sometimes no. Match the model to the task.

Claude 4 Model Family (Current)

All Claude 4 models support:

• Extended thinking
• Vision (images)
• Tool use
• Priority Tier access

Opus 4.6 and Sonnet 4.6 additionally support:

• Adaptive thinking (dynamic reasoning depth)
• 1M token context window (beta, via context-1m-2025-08-07 header — long context pricing applies beyond 200K)

AWS Bedrock IDs: anthropic.claude-opus-4-6-v1, anthropic.claude-sonnet-4-6 GCP Vertex AI IDs: claude-opus-4-6, claude-sonnet-4-6

Model Tiers

When Opus 4.6 IS Worth It

• ✅ Architecture decisions — Multi-file refactoring, system design
• ✅ Novel problem-solving — No clear pattern to follow
• ✅ Complex reasoning chains — Many dependencies, edge cases
• ✅ Long context understanding — Large codebases, documentation
• ✅ Nuanced judgment — Taste, style, UX decisions
• ✅ Learning sessions — Bootstrap learning, skill development
• ✅ Meditation/self-actualization — Meta-cognitive operations
• ✅ Extended thinking tasks — Deep analysis requiring internal reasoning

When Opus 4.6 IS Overkill

• ❌ Simple file edits — Renaming, adding imports
• ❌ Boilerplate generation — CRUD, scaffolding
• ❌ Format conversion — JSON ↔ YAML, etc.
• ❌ Syntax fixes — Lint errors, typos
• ❌ Documentation updates — README badges, version bumps

How LLM Choice Affects Alex

Alex's Cognitive Power by Model

Opus 4.6:     [████████████████████] Full cognitive architecture + deep thinking
Sonnet 4.6:   [██████████████████░░] Most capabilities, excellent for coding
Haiku 4.5:    [██████████████░░░░░░] Solid baseline, fast responses

With Opus 4.6, Alex can:

• Maintain 7±2 working memory rules across long sessions
• Execute complex meditation protocols with extended thinking
• Perform genuine meta-cognitive reflection
• Handle multi-file architecture changes
• Learn new skills through bootstrap learning

With Sonnet 4.6, Alex gets:

• Excellent coding capabilities (recommended for most development)
• 1M context window (beta) for large codebases
• 64K max output tokens + adaptive thinking
• Good cost-to-capability ratio
• Extended thinking support

With Haiku 4.5, Alex has:

• Near-frontier intelligence at lowest cost
• Fastest response times
• Good for routine operations

Cost Optimization Strategy

Knowledge Cutoffs

Auto Model Selection ⚠️

When using Auto in VS Code Copilot, the model switches dynamically based on task complexity. Alex cannot detect which model is currently running.

Tasks That REQUIRE Opus 4.5 (Warn User)

Warning Protocol

When user requests an Opus-level task while potentially on Auto/lesser model:

⚠️ Model Check: This task works best with Claude Opus 4.6. If you're using Auto model selection, please manually select Opus from the model picker for optimal results. Continue anyway?

Safe for Any Model

• Simple file edits, formatting
• Documentation updates
• Quick Q&A
• Code review (Sonnet+ recommended)
• Bug fixes (depends on complexity)

Practical Guidance

When to Upgrade Model Mid-Session

If you notice:

• Repeated mistakes on the same issue
• Losing context from earlier in conversation
• Superficial answers to complex questions
• Failure to see cross-file dependencies

→ Consider switching to a more capable model

When to Downgrade

If you're doing:

• Repetitive mechanical edits
• Simple Q&A
• Format conversions
• Quick lookups

→ Save cost with a faster model

The Alex Recommendation

For architecture evolution and complex cognitive tasks: → Always use Opus 4.6 — The cognitive architecture demands full capability

For production deployment, user-facing work: → Default to Sonnet 4.6 — Best balance of capability and cost → Allow Opus for complex tasks — User can request escalation

Token Economics