Claude Fable 5 with Aider

Setup

Steps one and two get you running. Steps three to five stop you from flying blind.

1. 1

   Install Aider and export your key

   Install with python -m pip install aider-install followed by aider-install, then export ANTHROPIC_API_KEY=your-key on macOS and Linux, or setx ANTHROPIC_API_KEY your-key on Windows and restart the shell. Aider also accepts the key through the --anthropic-api-key switch.

2. 2

   Run it once to confirm the route

   cd into your repo and run aider --model claude-fable-5. Aider recognises the ID and routes through the Anthropic API. Inside a session you can switch with /model claude-fable-5, which is the fastest way to escalate a single hard question without restarting.

3. 3

   Make it the default in .aider.conf.yml

   Create .aider.conf.yml in your git root or home directory with model: claude-fable-5 and cache-prompts: true. Aider checks the current directory first and falls back to the home file, so per-project overrides work: an expensive project can default to Fable 5 while everything else stays on a cheaper model.

4. 4

   Supply model metadata so costs are real

   Create .aider.model.metadata.json with a fully qualified key — anthropic/claude-fable-5, matching the litellm_provider field — and fill in litellm_provider, max_input_tokens of 1000000, max_output_tokens, mode: chat, input_cost_per_token of 0.00001 and output_cost_per_token of 0.00005. Without this, /tokens and the running cost display are meaningless.

5. 5

   Fix the parameters in .aider.model.settings.yml

   Add an entry for the model with use_temperature: false and cache_control: true. Do not copy an accepts_settings block containing thinking_tokens from the Claude 3.7 or Opus 4 examples in Aider's docs, and do not put a thinking budget under extra_params — both produce a 400 on this model.

Gotchas

Aider's defaults were tuned for a model generation that accepted more parameters than this one.

• Unknown metadata means Aider thinks the model is free

  Aider's own documentation says the unknown-context-window warning is usually safe to ignore, and for most models it is. It is not safe here. Aider will assume an unlimited window and zero cost, which means no context-limit protection on a 1M-token model and a cost readout of zero on the most expensive model you can run.

• map-tokens quietly sets your per-turn floor

  The repo map is prepended to requests, so map-tokens: 4096 means roughly four thousand tokens of input on every single turn before you have said anything. At Fable 5 input pricing that is a real number over a long session. Cache prompts to soften it, and drop the map size on repos where it is not earning its place.

• Aider sends temperature unless you tell it not to

  use_temperature is a per-model setting in Aider, and models it does not recognise do not get the false default that the Claude entries in its shipped settings carry. Fable 5 rejects temperature, top_p and top_k outright, so an unrecognised model plus a default temperature is a 400 on the first message.

• --thinking-tokens is a 400 on this model

  Aider's documented way to enable thinking is a budget_tokens value, either through the switch or through accepts_settings in the model settings file. Fable 5 removed budget_tokens: thinking is always on and cannot be configured that way. There is no equivalent Aider flag for effort, so depth is whatever the model defaults to.

Related guides