Verified August 2026
Claude Fable 5 in Claude Code
Claude Code is Anthropic's own coding agent, so it is the one client that accepted claude-fable-5 from day one. Here is how to select it, what to check, and the few ways it can still route you to a different model.
The short answer
Yes — Claude Code fully supports Fable 5. Since the Fable 5 carve-out, the model is selectable in Claude Code like any other model ID: open the model picker with /model and choose claude-fable-5, or set ANTHROPIC_MODEL / the model field in settings.json. It runs at default effort out of the box. The main thing to watch is that on safety-classifier fallback the client can silently route a turn to Opus 4.8, so verify the model indicator when it matters.
Support status
Yes — supported, verified August 2026. Claude Code is Anthropic's official coding agent, and claude-fable-5 is selectable in it like any other model ID since the Fable 5 carve-out. Because it is a first-party client it does not gate the model behind the allowlists that broke Zed's ACP bridge or older third-party editors. The one caveat is the safety-classifier fallback: on a flagged request Claude Code may serve the turn from Opus 4.8 instead, and the model indicator is the only reliable way to see it.
Setup
Five steps, most of them checking rather than installing.
- 1
Install or update Claude Code
Install or update to the latest Claude Code with npm install -g @anthropic-ai/claude-code, or use the native installer from Anthropic's docs. The model selector ships with the CLI, so being current avoids stale-model issues. Verify with claude --version.
- 2
Sign in
Run claude and complete authentication. A Claude Pro or Max subscription works, or you can bring your own Anthropic API key. Either is fine — unlike some editors, the subscription path is not the reason Fable 5 is missing.
- 3
Select claude-fable-5
Open the model picker with /model and choose claude-fable-5, or set it ahead of time via ANTHROPIC_MODEL in your environment or the model field in settings.json. The ID must be spelled exactly claude-fable-5 (hyphenated) or the picker shows nothing.
- 4
Set effort level
Claude Code exposes the effort parameter in settings. Fable 5 is normally run at the default effort; the high and max settings are where this site's benchmarks are measured. Because effort is a per-request API concept, you configure it once in Claude Code settings rather than per prompt.
- 5
Verify it is really routing to Fable 5
Check the model indicator in the Claude Code panel at the start of a session. The client can silently fall back to Opus 4.8 on a safety-classifier hit, so the picker saying Fable 5 is not proof every turn runs on Fable 5. If routing matters, confirm against the indicator (or usage data) rather than the selection.
Gotchas
Claude Code is the smoothest Fable 5 client, but it still has edges.
Effort interacts with planning mode
Effort is a per-request API parameter, but Claude Code configures it per session in settings. Running in planning mode can change how the model is invoked, and the effort you set is not always the effort the request actually gets. If you are measuring against the site's benchmarks (taken at high/max), match the setting before you compare.
Silent fallback to Opus 4.8
When a request trips the safety classifier, Claude Code can route the turn to Opus 4.8 instead of Fable 5 — and the model indicator is the only reliable way to notice. Anthropic reports this on under 5% of conversations on average. It is not a bug; it is the safeguard classifier deciding the request is safer handled by the older model.
The model ID must be exact
claude-fable-5 is hyphenated. A typo — claude_fable_5, claudefable5, claude-fable-5 with a trailing space — makes the picker show nothing or reject the selection. Copy the ID exactly, and if /model returns nothing, re-check the spelling before assuming the model is missing.
30-day retention applies to everything you send it
Fable 5 carries a mandatory 30-day retention window, and zero-data-retention agreements do not apply. Anything you send Claude Code on Fable 5 — including code and prompts — falls under that retention. If your org requires ZDR, do not route sensitive work to Fable 5.
Cost notes
Claude Fable 5 is $10 per million input tokens and $50 per million output, with a 1M context window and 128K max output (as of August 2026). Claude Code sessions are long-lived and token-hungry — a big refactor can chew through input tokens on context you re-send. Use the cost calculator to estimate a session before you start, and remember a safeguard fallback to Opus 4.8 bills at Opus 4.8 rates.
Estimate what a Claude Code session will cost
Claude Code sessions accumulate tokens fast. The cost calculator models Fable 5 pricing, effort, caching and the Opus 4.8 fallback. Or scan the common errors you may hit while running the CLI.