Aider with Claude through a gateway: ANTHROPIC_API_BASE, config files, architect mode and fixes

Aider sends every model call through LiteLLM, so pointing it at a gateway is a LiteLLM setting rather than an Aider one. For Claude that means ANTHROPIC_API_BASE set to the gateway's host root and a model name with the anthropic/ prefix. Two less obvious settings matter as much: newer Claude IDs need a model-settings entry, or Aider falls back to whole-file edits without a repo map.

Install Aider

python -m pip install aider-install
aider-install

# or the one-line installer (installs Python 3.12 if needed)
curl -LsSf https://aider.chat/install.sh | sh

powershell -ExecutionPolicy ByPass -c "irm https://aider.chat/install.ps1 | iex"

If the aider command isn't found afterwards, which is common on Windows, python -m aider runs the same program.

Connect to Claude through the gateway

export ANTHROPIC_API_KEY=your_gateway_key
export ANTHROPIC_API_BASE=https://aiprimetech.io

aider --model anthropic/claude-sonnet-5

setx ANTHROPIC_API_KEY your_gateway_key
setx ANTHROPIC_API_BASE https://aiprimetech.io
# open a new terminal, then:
aider --model anthropic/claude-sonnet-5

Aider's own docs don't cover a custom Anthropic base URL. The variable is read by LiteLLM, which checks ANTHROPIC_API_BASE first and then ANTHROPIC_BASE_URL, the variable Claude Code uses. Either works; if both are set, ANTHROPIC_API_BASE wins.

Keep the anthropic/ prefix on the model name. Without a provider prefix LiteLLM doesn't know where to send an unfamiliar name and fails with LLM Provider NOT provided.

Put the settings in config files

Aider reads .aider.conf.yml and .env from your home directory, the git repository root and the current directory, in that order, with later files taking priority. A file in the home directory covers every project.

model: anthropic/claude-sonnet-5
weak-model: anthropic/claude-haiku-4-5
set-env:
  - ANTHROPIC_API_BASE=https://aiprimetech.io
model-settings-file: ~/.aider.model.settings.yml
model-metadata-file: ~/.aider.model.metadata.json

ANTHROPIC_API_KEY=your_gateway_key

weak-model handles commit messages and chat-history summaries, so a cheap model there saves money on every session. set-env copies the variable into the environment before any model call, which keeps the base URL out of your shell profile.

Keep diff edits and the repo map on newer models

Aider picks its edit format by model name. Names it recognises, including those containing sonnet-4-, opus-4- or haiku-4-, get diff edits and the repository map. A name it doesn't recognise, such as claude-sonnet-5, falls back to rewriting whole files and drops the repo map. That makes edits slower and more expensive. A settings entry restores the right behaviour:

- name: anthropic/claude-sonnet-5
  edit_format: diff
  use_repo_map: true
  weak_model_name: anthropic/claude-haiku-4-5

- name: anthropic/claude-opus-5-5
  edit_format: diff
  use_repo_map: true
  weak_model_name: anthropic/claude-haiku-4-5
  editor_model_name: anthropic/claude-sonnet-5
  editor_edit_format: editor-diff

Unknown models also trigger a warning: Unknown context window size and costs, using sane defaults. It's harmless, but Aider then assumes an unlimited window and a zero price. The metadata file gives it real numbers:

{
  "anthropic/claude-sonnet-5": {
    "max_tokens": 64000,
    "max_input_tokens": 200000,
    "max_output_tokens": 64000,
    "input_cost_per_token": 0.000002,
    "output_cost_per_token": 0.00001,
    "litellm_provider": "anthropic",
    "mode": "chat"
  },
  "anthropic/claude-opus-5-5": {
    "max_tokens": 64000,
    "max_input_tokens": 200000,
    "max_output_tokens": 64000,
    "input_cost_per_token": 0.000004,
    "output_cost_per_token": 0.00002,
    "litellm_provider": "anthropic",
    "mode": "chat"
  }
}

The prices are Anthropic's list prices per token, so Aider's cost readout shows list cost; the gateway bills less than that. A 200K input limit is a safe working ceiling even where a model supports more.

Architect mode: plan with one model, edit with another

aider --architect --model anthropic/claude-opus-5-5 --editor-model anthropic/claude-sonnet-5

In architect mode the main model proposes the change and the editor model turns it into file edits. A stronger model for planning and a cheaper one for editing usually costs less than running the strong model for both. Inside a session, /architect switches to it for one message and /chat-mode architect switches for the rest of the session.

The OpenAI-compatible route

The gateway also speaks the OpenAI Chat Completions format, which is the route for GPT models on the same account. For this route the base URL does include /v1, and the model name takes the openai/ prefix:

export OPENAI_API_KEY=your_gateway_key
export OPENAI_API_BASE=https://aiprimetech.io/v1

aider --model openai/claude-sonnet-5

For Claude, prefer the native route above: model metadata and prompt caching line up with Anthropic's format. If a stray OPENAI_BASE_URL is set in your environment, LiteLLM uses it in preference to OPENAI_API_BASE.

Models

Aider's short aliases (sonnet, opus, haiku) point at older model IDs, so name the model explicitly. Old dated IDs such as claude-3-5-sonnet-20241022 are not served.

Troubleshooting

For errors that come from the model side rather than from Aider, see the Claude API errors guide.

Frequently asked questions

AI Prime Tech is an independent API gateway. It is not affiliated with, endorsed by, or a reseller of Anthropic. Claude and related model names are trademarks of their respective owners.