Model Configuration & Environment Variables

Model Aliases

Alias	Behavior
default	Clears overrides; recommended model for your account
best	Fable 5 if your org has access, else latest Opus
fable	Claude Fable 5 — hardest, longest-running tasks
opus	Latest Opus (Opus 4.8 on Anthropic API) — complex reasoning
sonnet	Latest Sonnet (Sonnet 5) — daily coding default
haiku	Fast/cheap — simple tasks
sonnet[1m] / opus[1m]	1M-token context window variants
opusplan	Opus during plan mode, Sonnet for execution

Aliases track the newest version. Pin exact versions with full names (claude-opus-4-8) or ANTHROPIC_DEFAULT_OPUS_MODEL etc.

Fable 5 notes

• Most capable model; excels at long autonomous sessions, root-cause investigation, verification
• Not the default — select with /model fable
• Give it outcomes, not steps; hand it ambiguous problems; skip “remember to test” reminders

Setting the Model (priority order)

1. /model <alias|name> during session — saves as default for new sessions (s in the picker = this session only). Switching mid-session invalidates the prompt cache; the picker warns.
2. claude --model <alias|name> at startup (session-only)
3. ANTHROPIC_MODEL=<alias|name> env var (session-only)
4. "model": "opus" in settings.json (permanent)

Resumed sessions keep the model they were saved with.

Effort & Fast Mode

• /effort low|medium|high — reasoning depth vs speed/cost; inherited by subagents
• /fast — fast mode (Opus with faster output; billed at a premium). Great for iteration loops, wrong for hard design work

Key Environment Variables

Variable	Purpose
ANTHROPIC_MODEL	Model for this session
ANTHROPIC_DEFAULT_OPUS_MODEL / ..._SONNET_MODEL / ..._HAIKU_MODEL / ..._FABLE_MODEL	Pin what each alias resolves to
CLAUDE_CODE_SUBAGENT_MODEL	Default model for subagents
ANTHROPIC_API_KEY	API-key auth
ANTHROPIC_BASE_URL	Route via LLM gateway (changes where, not which model)
CLAUDE_CONFIG_DIR	Move ~/.claude storage elsewhere
CLAUDE_CODE_SKIP_PROMPT_HISTORY	Don’t write transcripts
ENABLE_TOOL_SEARCH	auto/false — deferred vs upfront MCP tool schemas
CLAUDE_CODE_SUBPROCESS_ENV_SCRUB	Strip credentials from subprocesses
CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS	Enable agent teams (see 32-agent-teams.md)

Set env vars persistently via settings.json:

{
  "model": "sonnet",
  "env": {
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "claude-sonnet-5",
    "ENABLE_TOOL_SEARCH": "auto"
  }
}

Enterprise Controls (awareness level)

• availableModels — allowlist of selectable models (managed settings)
• enforceAvailableModels — make the “Default” picker row obey the allowlist
• Org admins (Enterprise) can also disable models / set org default from the admin console

Real Case: Model Strategy for a Team

// .claude/settings.json (project, committed)
{
  "model": "sonnet"
}

• Daily work: Sonnet (fast, cheap, excellent at coding)
• Design sessions: /model opusplan — Opus plans, Sonnet executes
• Nightly deep bug hunt (routine): fable
• CI helpers / commit-message generation: haiku via claude -p --model haiku

🧠 Quick Recall

• Trick: “fable hardest, opus reasons, sonnet ships, haiku sprints.”
• opusplan = brains plan, hands type.

👨‍🏫 Teach It

• Show: /model picker — Enter saves default, s = this session only.
• They’ll trip on: switching models mid-task — it wipes the prompt cache (slow + expensive turn).

---

Learning path: ⬅ 12-settings-system.md · Index · ➡ 14-permissions.md