Verified July 27, 2026
Claude Fable 5 in Cline
Enabling it takes thirty seconds. Deciding which turns deserve a $50-per-million-output model is the actual work.
The short answer
Supported, no workaround needed. Pick the Anthropic provider with your own key, or Cline's usage billing, and select Claude Fable 5. Then immediately turn on 'Use different models for Plan and Act' — Fable 5 for planning, Sonnet for execution — because in Cline the cost driver is the file context resent on every turn, not the model's answer.
Support status
Fully supported, verified July 27, 2026. Cline documents Claude Fable 5 on its own model page and offers two billing surfaces: the Cline (usage-billing) provider and the Anthropic provider with your own API key. Cline's own docs flag the same two caveats Anthropic does — Fable is a higher-cost, slower tier than the speed-oriented models, and its safety classifiers can decline requests, so any production workflow needs a defined fallback. The one real limitation is that the Anthropic provider's model list is curated by the Cline team, so arbitrary model IDs cannot be typed in.
Setup
Four steps, and the third one is the one that saves you money.
- 1
Choose a billing surface
The Cline (usage-billing) provider means no API key to manage and one invoice; the Anthropic provider means your own key, your own rate limits, and your usage visible in the Anthropic console alongside everything else you run. For teams already tracking Anthropic spend centrally, the second is easier to attribute. Neither changes the model's behaviour.
- 2
Select the provider and model
Open Cline settings with the gear icon, choose Anthropic in the API Provider dropdown, paste your key into the Anthropic API Key field, and pick Claude Fable 5 from the Model dropdown. Leave the custom base URL checkbox alone unless you are deliberately routing through a proxy.
- 3
Split Plan and Act across two models
In Cline settings, enable 'Use different models for Plan and Act', then assign Fable 5 to Plan and a cheaper model to Act. Cline's docs list this as a standard cost-optimisation pattern. It fits Fable 5 unusually well: planning is where deep reasoning over a large repository pays off, while Act mode is mostly applying a decision that has already been made.
- 4
If you route through a gateway, namespace the ID
Gateways expose models as provider/model, so a bare claude-fable-5 fails and you need anthropic/claude-fable-5. Since the Anthropic provider only accepts IDs from Cline's curated list, gateway users go through the LiteLLM provider or the OpenAI Compatible provider, both of which take a custom base URL and a free-form model ID.
Gotchas
Three things that surprise people in week one.
The 1M window is a spending permission, not a feature
Cline resends conversation and file context on every turn. With a 1M window nothing forces a truncation, so a long task quietly grows its per-turn input until each step costs several dollars in input alone. Break long work into explicit checkpoints and start fresh tasks rather than letting one thread run all afternoon.
You cannot type a custom model ID
Cline's Anthropic provider accepts only model IDs from a list the team maintains — this is documented behaviour, discussed at length in cline/cline issue #4076, not a bug you can configure around. If you need an ID the list does not have, the maintainers' own recommendation is a LiteLLM proxy in front, with Cline pointed at the proxy.
Refusals are part of the deal
Cline's own model page lists this as a limitation: Fable 5's safety classifiers can decline requests, and production workflows should handle refusals and define a fallback. In an autonomous agent loop this matters more than in chat, because a refusal mid-task leaves the workspace half-modified. Decide in advance which model picks up the thread.
Cost notes
Fable 5 is $10 per million input and $50 per million output, against $3 and $15 for Sonnet 5 — roughly 3.3x on both sides. But that ratio understates the difference in Cline, because an agent loop sends far more input than it receives output: file contents, diffs and tool results all go up, a few hundred tokens of decision come back. Prompt caching returns 90% of the input price on cache hits, so the highest-leverage moves are keeping the cache warm (don't churn models mid-task), splitting Plan and Act, and closing tasks when they are done.
Know what a turn costs before you run it
Our pricing breakdown shows the input/output/cache math for Fable 5 side by side with Opus 4.8 and Sonnet 5, including the cached-read discount that changes the picture for agent loops.