# Claude model selection checklist

Checked 7 October 2026

Use this worksheet for one workflow. A model that wins here has won this job, not every job.

## 1. Job

Accepted output:

Person responsible:

Failure impact:

Examples in the test set:

## 2. Source and data

Source of truth:

Fields, files or date range required:

Data that must not leave the current environment:

Required processing region or contract:

## 3. Surface

Choose the simplest viable route: Claude chat or Cowork, Claude Code, approved connector or MCP, Claude API, or Managed Agent.

Why this surface is needed:

## 4. Model shortlist

Sonnet 5.5 for well-scoped production work:

Opus 5.5 for difficult judgement or architecture:

Fable 5.1 only if the long-horizon job still fails with Opus:

Haiku 4.5 for fast, lower-complexity volume:

## 5. Tools and permissions

Read tools allowed:

Draft or write tools allowed:

Always ask before:

Always deny:

## 6. Acceptance test

Pass conditions:

Edge cases:

Deterministic checks:

Human review rule:

## 7. Cost and limits

Maximum task cost:

Maximum turns or runtime:

Latency requirement:

Fallback trigger:

## 8. Release decision

Accepted result rate:

Material error rate:

Reviewer time:

Chosen route and reason:

Next model review date: