Aider Costs: The Free Client and the Actual API Charges You Pay

Aider, the open-source AI pair-programming tool for the terminal, costs nothing to use. It is Apache-2.0 software, and aider.chat has no plans, tiers, seats, or hosted service. So “Aider costs” effectively means the model API charges incurred through Aider. For budgeting, look at tokens per session × the provider's rate. The setting most likely to inflate that amount is Aider's default behavior when it encounters an unknown model ID. It takes two lines to fix, as explained below.

First, a distinction. Search results also include another company called Aider that sells accounting software for firms. Its published monthly prices are for that product and have nothing to do with this CLI. This page covers only the CLI at aider.chat, the tool distributed as the aider-chat package from the Aider-AI/aider repository.

What you pay Aider, and what you are actually billed for

Aider's FAQ says the tool is released as Apache 2.0 open-source software, and the homepage, FAQ, installation page, and LLM connection page contain no prices, plans, or subscription language (verified October 1, 2026). All billing happens on the other side of the endpoint.

Check the update status before standardizing on it for a team. The repository is not archived and has 49,309 stars, but main had its last commit on May 22, 2026 (GitHub API, October 1, 2026). The official “latest version” differs between two sources: PyPI lists 0.86.2, published February 12, 2026, while GitHub Releases stops at v0.86.0, dated August 9, 2025. Installing with pip install gets the PyPI version. For the Python version requirement, pip actually enforces >=3.10,<3.13 from the 0.86.2 package metadata rather than the description on the installation page.

The setting that most affects API costs

This is the part you won't see on a price list. Aider chooses the edit format, repository map (repo map), and caching behavior by matching the model ID against a built-in table. The lookup uses the entire string after --model, including its prefix. If nothing matches, the defaults are edit_format: whole, use_repo_map: false, cache_control: false, and weak_model_name: null, according to the advanced model settings page.

In the source for release 0.86.2, matching happens in two steps. First it checks for an exact match among 313 full names, such as gemini/gemini-3-pro-preview and openai/gpt-5.2. Only if none match does it try substring rules, including sonnet-4-5, haiku-4-5, and gpt-4.1. For an OpenAI-compatible gateway, the model ID must be passed as openai/<id>. The 40 entries in the table with a openai/ prefix are all OpenAI's own model names. So openai/claude-sonnet-4-6 matches neither step, and runs with whole-file rewrites and the repository map disabled. The model outputs each edited file in full on every turn, and output tokens are the more expensive part of the bill.

Two lines fix this. Aider reads this file from the home directory, Git root, or current directory. The key is the full name passed to --model.

# name은 --model에 넘기는 문자열과 정확히 같아야 합니다(openai/ 접두사 포함)
- name: openai/claude-sonnet-4-6
  edit_format: diff
  use_repo_map: true

For a one-time run, you can also add --edit-format diff at launch. Conversely, openai/claude-haiku-4-5 matches the substring haiku-4-5, which automatically enables diff format and the repository map. That is a coincidence in the name and could change with any Aider version, so verify it yourself rather than relying on it.

One more easily overlooked effect: Aider calls the API in three roles—the main model, an editor model for --architect mode, and a weak model for commit messages and conversation-history summaries (option reference). When the ID is not recognized, the source sets both weak and editor to the main model, so all three roles are billed at the most expensive rate. Avoid this by specifying a cheaper ID with --weak-model (the openai/ prefix is required here too).

Finally, an unrecognized ID also makes cost reporting disappear. Without input_cost_per_token, Aider prints only token counts and no cost line; the warning page says it “assumes the model is free.” The following metadata file restores the cost display.

{
  "openai/claude-sonnet-4-6": {
    "input_cost_per_token": 0.0000021000,
    "output_cost_per_token": 0.0000105000,
    "litellm_provider": "openai",
    "mode": "chat"
  }
}

Even with metadata, Aider's amount is its own calculation based on reported tokens, and the cache multipliers are hard-coded for specific providers. Treat it as a reference and confirm against the provider's usage records.

Estimated cost for one session

The following is token arithmetic, not a measured task cost or a billing cap. It assumes one session sends 200,000 uncached input tokens and receives 12,000 output tokens using diff edits. The whole column gives an illustrative figure assuming the same task produces 40,000 output tokens because entire files are output again; it is not a measurement. Rates are current prices per million tokens from the Kunavo catalog.

Under this assumption, the editing format alone changes the cost of the same Claude Sonnet 4.6 task from $0.546 to $0.840. That is why you should correct the configuration before comparing provider rates. If commit messages and summary traffic (assuming 30,000 input and 2,000 output tokens) remain on the main model, Claude Sonnet 4.6 costs $0.084; assigning the weak model to Claude Haiku 4.5 brings the cost to $0.028.

Multiply by your number of sessions per day to turn this into a budget. Kunavo catalog amounts are the billing floor, not a ceiling. When the upstream provider reports its costs, you are charged the higher of the catalog amount or upstream cost × the applicable margin. Cache fees and external tools are not included in this example.

“Cheapest API” is two different questions

Aider has only one benchmark for task costs that can be independently checked. Its polyglot benchmark has each model solve the same 225 Exercism problems across 6 languages under the same conditions, and records the cost.

These figures come from the leaderboard data file. Two caveats matter more than the ranking: the latest entry is dated October 3, 2025, so models released afterward have never been measured here. The cost column has no price list or date, so it reflects the cost of that run at the time, not a rate available today. These models are cited as benchmark evidence; DeepSeek is not in the Kunavo catalog. The lower-cost option in the estimate table above is Claude Haiku 4.5.

The practical takeaway is this: o3-pro cost five times as much as gpt-5 (high) and scored lower, so a higher rate does not guarantee completion. A lower rate does not guarantee the lowest total either; a cheap model that needs three retries can cost more than an expensive model that finishes in one attempt. Choose the least expensive model that completes the task with an acceptable review burden, then verify using the actual amount recorded in your account for that task. Aider's /tokens is, as defined in the command reference, the token count in the current chat context—not the total billed so far.

Which route fits which use case

Two limits of the gateway route. Gateways without an Aider-specific module connect using OPENAI_API_BASE, OPENAI_API_KEY, and the required openai/ model ID prefix, as described on the OpenAI-compatible API page. The documentation does not describe a setting to override the Anthropic-format base URL, and the string ANTHROPIC_API_BASE does not appear in the 0.86.2 source. Do not expect prompt-cache discounts on this route either. The caching page limits supported providers, requires --cache-prompts to be explicitly enabled, and says cache statistics and costs are unavailable during streaming. This page has not tested whether caching works on the compatible route.

Using Aider with Kunavo, and how to pay from Korea

Kunavo has a setup guide for this client. It is a public configuration reference, not a compatibility test. We have not run Aider against the Kunavo endpoint; everything on this page comes from Aider's released source and official documentation. Keep your current route, run one well-scoped task, and then check the amount recorded in your account.

Payment is prepaid top-up, with no monthly fee. The minimum top-up is $10, the minimum amount needed to create a balance rather than a task fee or subscription. Payment uses Stripe Checkout; in Korea, you can use cards (Visa, Mastercard, American Express, JCB, UnionPay), Apple Pay, Google Pay, and Link, and when the payment page is displayed in KRW, Kakao Pay, Naver Pay, PAYCO, Samsung Pay, and domestic cards (including cards that do not support international payments) also appear. The KRW conversion includes a 2–4% currency-conversion fee paid by the buyer, which does not apply when paying in dollars (Stripe Adaptive Pricing documentation, checked 2026-10-03). Toss is not supported. These methods top up the Kunavo balance used for Aider; they are not direct payments to Anthropic or OpenAI. For details, see payment information, and when you are ready, create an account. The more detailed English original is Aider pricing.

Frequently asked questions

Aider's repository, package, documentation, and leaderboard data were checked on October 1, 2026. Descriptions of behavior are based on the source for the official v0.86.2 release. Kunavo rates are taken from the live catalog, and all dollar examples on this page are illustrative token calculations, not measured task costs.