Aider: the openai/ prefix and model settings files

Aider is a terminal pair programmer that works directly on your git repo. Because it goes through LiteLLM, a custom endpoint needs the openai/ prefix in front of the model name. This page was checked against the docs for version 0.86.

Connecting

export OPENAI_API_BASE="https://api.uttapen.ir/v1"
export OPENAI_API_KEY="sk-up-..."

aider --model openai/anthropic/claude-sonnet-5

The openai/ prefix tells Aider which protocol to speak; the rest of the string stays exactly as our model id, because only the first / is split off. So anthropic/claude-sonnet-5 becomes openai/anthropic/claude-sonnet-5, and openai/gpt-5-mini becomes openai/openai/gpt-5-mini — yes, openai twice.

Instead of environment variables you can keep it all in .aider.conf.yml, in your home directory or the repo root:

openai-api-base: https://api.uttapen.ir/v1
model: openai/anthropic/claude-sonnet-5
weak-model: openai/openai/gpt-5-nano
edit-format: diff

Keep the key out of that file; pass OPENAI_API_KEY through the environment.

Output ceiling and model metadata

Aider sends no max_tokens of its own. On our side that means the hold is based on a default estimate, and if the model writes a long answer the real amount is still deducted in full, so the balance can briefly go negative (pricing). Set the ceiling in .aider.model.settings.yml:

- name: openai/anthropic/claude-sonnet-5
  edit_format: diff
  use_repo_map: true
  extra_params:
    max_tokens: 8192

And so Aider knows the context size (and stops warning about an unknown context window), add .aider.model.metadata.json next to it:

{
  "openai/anthropic/claude-sonnet-5": {
    "max_input_tokens": 200000,
    "max_output_tokens": 8192,
    "input_cost_per_token": 0,
    "output_cost_per_token": 0,
    "litellm_provider": "openai",
    "mode": "chat"
  }
}

Both files are read from your home directory, the repo root or the current folder. Each key must be the full Aider model name, prefix included; matching is exact. The cost fields are zero here so you don't get a misleading dollar estimate — the real toman figure is in Dashboard → Usage.

Set edit_format: diff and use_repo_map: true too. For names it doesn't recognise, Aider falls back to rewriting whole files with no repo map, which is both weaker and more expensive. If a model rejects temperature, add use_temperature: false to the same entry.

A weak model for small jobs

Aider hands commit messages and chat-history summaries to the weak model, and reuses the main model when you don't set one. Give it something cheap:

aider --model openai/anthropic/claude-sonnet-5 \
      --weak-model openai/openai/gpt-5-nano

For cheaper everyday work, qwen/qwen3-coder and deepseek/deepseek-v3.2 make good main models.

Checking that it works

aider --model openai/openai/gpt-5-nano --message "hello" --no-auto-commits

If you get an answer, the same request should appear in Dashboard → Usage.

Troubleshooting

If your available balance doesn't cover the whole max_tokens, uttapen lowers the output ceiling to fit it (floor 1,024 tokens) and returns the granted value in the X-Uttapen-Max-Tokens header. The diff can then arrive truncated and Aider won't be able to apply it; topping up fixes that. 402 is reserved for the case where even the floor doesn't fit.

Other tools are in integrations; error codes are in errors.