select-model — Choose the right Claude model

Determine which Claude model (Fable 5, Haiku 4.5, Sonnet 4.6, or Opus 4.8) is best for your task. Runs in procedural mode (decision tree reference) or executable mode (analyze a task and recommend a model with optional config update).

Procedure

Model tiers at a glance

Fable 5 / Opus 4.8 (Tier 1: Conductor & Deep Reasoning) - Highest reasoning depth, strategic planning, adversarial review - Handles ambiguous requirements, complex architectural refactors, multi-step dependencies - Best for: Orchestrator session conductor, adversarial self-review, subtle debugging, security analysis - Strategy: Use for high-leverage judgment, planning, and final auditing

Sonnet 4.6 / 5 (Tier 2: High-Velocity Execution) - Balanced: strong coding capability, fast execution, cost-effective - Handles task execution, test generation, refactoring, boilerplate - Best for: Subagent worker execution, feature implementation from clear specs, documentation - Strategy: Delegate bounded implementation sub-tasks to maximize speed and token efficiency

Haiku 4.5 (Tier 3: Fast Verification & Scans) - Speed + lightweight focused reasoning - Best for: Simple queries, single-file regex/syntax checks, triage, shallow unit tests - Limitations: Not suited for cross-file architecture or deep review

Orchestrator-Worker Delegation Pattern

When orchestrating complex or multi-step work: 1. Conductor (Tier 1): Breaks down the objective into modular, bounded sub-tasks with strict checklists and unambiguous acceptance criteria. 2. Workers (Tier 2/3): Subagents execute bounded tasks with tight stop conditions (preventing open-ended token burns or side quests). 3. Verification (Tier 1): Conductor audits worker diffs and runs adversarial review before committing.

Decision tree: Pick your model

  1. Is the task trivial or a fast scan? (single query, shallow check)
    • YES → Haiku 4.5 (save cost)
    • NO → Continue
  2. Is the task bounded implementation from a clear spec? (writing code, generating tests, refactoring)
    • YES → Sonnet 4.6 / 5 (high-velocity executor)
    • NO → Continue
  3. Does the task require high-level orchestration, deep design, or adversarial review? (architectural decisions, subtle bugs, security audit, orchestrator loop)
    • YES → Fable 5 / Opus 4.8 (most capable)

Task → Model mapping

Task Category Complexity Recommended Tier Rationale
Simple query / scan ⭐ Tier 3 (Haiku 4.5) Minimal reasoning; speed matters
Code snippet / boilerplate ⭐ Tier 2 (Sonnet 4.6) Fast and accurate generation
Bug fix (clear pattern) ⭐ Tier 2 (Sonnet 4.6) Clear problem, straightforward fix
Refactor (bounded) ⭐⭐ Tier 2 (Sonnet 4.6) Follows established architecture
Subagent execution ⭐⭐ Tier 2 (Sonnet 4.6) High throughput, token efficient
Multi-file architecture ⭐⭐⭐ Tier 1 (Fable 5 / Opus 4.8) Needs high reasoning; conductor tier
Subtle bug hunt ⭐⭐⭐ Tier 1 (Fable 5 / Opus 4.8) Deep lateral reasoning required
Adversarial review ⭐⭐⭐ Tier 1 (Fable 5 / Opus 4.8) Strict verification against subtle issues
Orchestration & planning ⭐⭐⭐ Tier 1 (Fable 5 / Opus 4.8) Decomposes task DAGs with stop criteria

Model selection factors

  • Reasoning depth: Tier 3 < Tier 2 < Tier 1
  • Code generation: Tier 2 provides optimal throughput/cost balance; Tier 1 handles architectural nuance
  • Code review: Tier 1 recommended for adversarial passes
  • Token efficiency: Use Tier 1 for judgment and Tier 2/3 for generation

Executable mode (auto-recommend and config update)

Instead of consulting the decision tree manually, invoke the skill with a task description. The script will analyze complexity, check your current settings, recommend the right model, and optionally suggest a config update:

/select-model --task "I need to refactor a critical payment module with security implications"

The script outputs a recommendation and optionally suggests updating ~/.claude/settings.json to use the recommended model for future sessions.

How to use

  • Procedural mode (manual decision tree): Read this procedure, follow the decision tree or task mapping, choose your model.
  • Executable mode (auto-recommend): Invoke /select-model --task "<task description>" and the script provides a personalized recommendation and config suggestion.
  • Chained from assess-model-fit: If /assess-model-fit recommends escalation, it auto-invokes select-model with your task details.

FAQ

Q: I’m using Haiku 4.5 but it keeps failing. What now? A: Escalate to Sonnet (Tier 2) or Fable/Opus (Tier 1). Use this skill to confirm the better tier for your task.

Q: Opus is expensive. Can I use Sonnet or Fable 5 instead? A: Yes. If your task doesn’t need deep architectural design, use Sonnet for execution. For orchestration or review, Fable 5 or Opus provides the needed reasoning depth.

Q: Should I always use the highest model? A: No—higher models are costlier. Match the model to task complexity. Use Haiku 4.5 for simple work, Sonnet for execution, and Fable 5 or Opus for orchestration and adversarial review.

Q: Can I pick a model manually without using this skill? A: Yes. This skill is advisory. You can always override and choose your own model.

Back to top