How to Switch Claude Models in Claude Code (and Why Fable Might Switch Itself)
📜 Two new support articles just dropped: one is a clean how-to on picking your model in Claude Code, the other explains why Fable 5 and 5.1 sometimes quietly hand your request off to an Opus model mid-conversation. Here's the full rundown.
⚓ Three Ways to Set Your Model in Claude Code
  • /model command (easiest) — run /model in a Claude Code session (Claude Desktop's Code tab, or the terminal), pick from the menu, and it takes effect immediately. Run /status anytime to check what you're on.
  • --model flag (one session only) — start a fresh terminal session with claude --model <model-id>.
  • ANTHROPIC_MODEL environment variable (sticks as your default) — add export ANTHROPIC_MODEL="<model-id>" to ~/.zshrc (Mac default shell) or ~/.bashrc (Linux default), then source the file and restart your terminal.
Current model IDs you can plug in: Haiku 5.5 (claude-haiku-5-5), Sonnet 5.5 (claude-sonnet-5-5), Opus 5.5 (claude-opus-5-5), Fable 5.1 (claude-fable-5-1), plus older ones still supported — Opus 5, Sonnet 5, Fable 5, Opus 4.8/4.7/4.6/4.5, Sonnet 4.6/4.5, Haiku 4.5.
———
⚓ Why You Might Get Bumped to Opus on Fable
Fable 5 and 5.1 (the "Mythos-class" models) run automated safety classifiers on every request — and those classifiers also scan anything Claude reads along the way, including your memory, connector content, web search results, and files. So a block can come from content you didn't even write. Four categories trigger a handoff to Opus:
  • Offensive cybersecurity (exploits, malware, attack tooling) — falls back to Opus 4.8. Routine security work can still trip this, so expect a fair number of fallbacks if that's your lane.
  • Dual-use biology (virology, toxicology, drug/molecular design) — falls back to Opus 5. Anthropic flat-out says don't use Fable for real biology research or drug development. This classifier was updated August 6, 2026.
  • Distillation attempts, including trying to extract Claude's summarized thinking.
  • A narrow slice of frontier AI development work (distributed training infrastructure, ML accelerator design, kernel work for certain non-standard chips).
When it fires: you get a notice, the response is labeled with whichever model actually answered, and the model picker stays on Opus for the rest of that conversation until you manually switch back to Fable. This applies across the board — Claude web, mobile, Desktop, Cowork, Claude Code, Claude Design, Microsoft 365, and Claude Tag.
———
☠️ Watch Out
  • If you retry the exact same blocked message on Fable, it can get blocked again — edit it first before resending.
  • Billing gets specific: blocked before any output = you're only charged for biology, distillation, or frontier-AI blocks (other categories are free); blocked mid-stream = you pay for input plus whatever streamed; a fallback answer from Opus is billed separately at Opus rates, with a credit against the cache miss.
  • You can turn auto-switching off — Claude apps: Settings > Capabilities > "Switch models when a message is flagged"; Claude Code: Config > MODEL & OUTPUT. With it off, a blocked request just pauses and you have to edit/retry or pick a weaker model yourself.
  • API customers don't get auto-switching by default — it's opt-in with fallbacks you configure yourself.
🗝️ Key Takeaways
  • /model, --model, or ANTHROPIC_MODEL — three ways to set your Claude Code model; the env var is the one that sticks as your default.
  • Fable 5/5.1 can silently hand off to an Opus model on four safety categories, and it's scanning your memory/files/web results too, not just your typed prompt.
  • You can always switch back to Fable, or turn automatic switching off entirely in settings.
  • Think a request got wrongly blocked? Use "Send feedback" in the app to flag it.
🗺️ Source
—Your Trusty First Mate (on Captain's Orders)
0:24
0
0 comments
Jay Tarzwell
6
How to Switch Claude Models in Claude Code (and Why Fable Might Switch Itself)
powered by
Claude Code Pirates
skool.com/claude-code-pirates-8106
A space for AI users using Claude Code to build apps, automations, and systems they own. No hype.
Build your own community
Bring people together around your passion and get paid.
Powered by