Verified July 27, 2026
Claude Fable 5 in Continue.dev
A config format built around token budgets, meeting a model that removed them.
The short answer
Supported for chat, edit and apply with a plain anthropic provider block — but with one caveat worth knowing before you commit. Continue's reasoning controls are reasoning and reasoningBudgetTokens, which target the legacy thinking budget that Fable 5 rejects, and we found no documented effort or output_config passthrough in the config.yaml reference. So Fable 5 runs at its default effort of high, and the only way down is a proxy in the middle.
Support status
Works with a caveat, verified July 27, 2026. Continue's Anthropic provider takes any model ID you write in config.yaml, so claude-fable-5 needs no special support to run — this is the opposite of Cline's curated list. The caveat is on the parameter side. Continue's documented Anthropic reasoning options are the boolean reasoning and the numeric reasoningBudgetTokens, both descended from the Claude 3.7 thinking-budget era. Fable 5 removed budget_tokens entirely and replaced it with output_config.effort, and passing the old shape returns HTTP 400. As of our verification date we found no effort or output_config field in Continue's config.yaml reference.
Setup
Four steps, and step two is mostly about what to leave out.
- 1
Open your config.yaml
The global config lives in ~/.continue on macOS and Linux and %USERPROFILE%\.continue on Windows. If a config.yaml exists it is loaded instead of the older config.json, so migrate rather than keeping both. The file needs a name, a version and schema: v1 at the top.
- 2
Add the model with provider anthropic
Under models, add an entry with name, provider: anthropic, model: claude-fable-5 and apiKey. Set roles to chat, edit and apply. Do not add temperature to defaultCompletionOptions and do not set reasoning or reasoningBudgetTokens — Fable 5 rejects sampling parameters and the legacy thinking budget with a 400.
- 3
Turn on prompt caching
Add defaultCompletionOptions with promptCaching: true. Continue documents this as the way to cache the system message and turn-by-turn conversation with Claude models. At Fable 5's input price this is the single highest-return line in the file, worth a 90% discount on cached reads.
- 4
Reload config and give autocomplete a different model
Click Reload config in the config selector in the Continue extension. Then add a second, small model with roles: autocomplete — a frontier reasoning model behind every keystroke is both slow and expensive, and Fable 5 always thinks, so there is no low-latency mode to fall back to.
Gotchas
All three come from the same root: Continue's Anthropic options predate this model generation.
Never give Fable 5 the autocomplete role
Continue's default roles are chat, edit, apply and summarize, and autocomplete has to be assigned deliberately — which is a good default. Assigning it to Fable 5 means firing an always-thinking, $10-per-million-input model on a debounce timer. Keep a small dedicated model for that role.
There is no effort field, so you pay the default
Fable 5's effort defaults to high, and high is the expensive end of a scale that runs low, medium, high, xhigh, max. Anthropic's own guidance is that medium and low still perform well on this model. Without a passthrough field, Continue users cannot take that discount from inside the editor — a LiteLLM or gateway layer that injects output_config is the workaround.
reasoningBudgetTokens returns a 400
Continue exposes reasoning and reasoningBudgetTokens for Claude 3.7 and later, which produce the thinking block with budget_tokens. On Fable 5 that parameter was removed: sending it, or sending thinking type enabled or disabled, returns HTTP 400. The error text names the parameter, so it is recognisable — but only if you know to look at your own config rather than the model.
Cost notes
Because effort is stuck at the default, Continue is the one integration on this site where you cannot buy a cheaper answer — every request is a high-effort request at $10 per million input and $50 per million output. That makes the two available levers matter more: promptCaching: true for the 90% cached-read discount, and role splitting so Fable 5 only holds the chat role while a cheaper model takes edit, apply and autocomplete. If effort control is a hard requirement, put LiteLLM in front and set output_config there.
Check whether Continue has caught up
The effort gap is a version problem, not a permanent one. Our status page records what we last confirmed in each tool, with the date we confirmed it.