Aider Pricing: Free Client and the Real Cost of the API

Aider, the open-source AI coding assistant that runs in the terminal, costs $0. It is Apache-2.0 software, and aider.chat has no plans, tiers, seats, or hosted service—so “Aider pricing” effectively means the API bill for the model you call through it. The amount to budget for is tokens per session multiplied by your provider's rate, and the setting most likely to inflate that number is Aider's default behavior when it encounters a model ID it does not recognize. It takes two lines to fix, as explained below.

A clarification first, because search results mix the two: another company, which makes accounting close software for firms, also uses the name Aider and publishes a monthly minimum. Those prices are for that product. This page covers only the CLI from aider.chat, distributed as the aider-chat package from the Aider-AI/aider repository.

What Aider charges and what actually charges you

The Aider FAQ says the tool is published as open source under Apache 2.0, and there is no pricing, plan, or subscription text on the homepage, FAQ, installation page, or LLM connection pages (verified on October 1, 2026). The entire bill is on the other side of the endpoint.

Check the release cadence before standardizing on the tool across a team. The repository is not archived and has 49,309 stars, but the last commit on main was on May 22, 2026 (GitHub API, October 1, 2026). The two official sources for the “latest version” also disagree: PyPI serves 0.86.2, published on February 12, 2026, while GitHub Releases stops at v0.86.0, from August 9, 2025. It is the PyPI version that pip install installs. For Python, the constraint in the 0.86.2 package metadata, >=3.10,<3.13, is what pip actually enforces.

The setting that affects API costs the most

This is the part that pricing tables don’t show. Aider chooses the edit format, repo map, and cache behavior by comparing the model ID with a built-in table—and the ID being compared is the entire string you type after --model, including the prefix. When nothing matches, the defaults are edit_format: whole, use_repo_map: false, cache_control: false, and weak_model_name: null, according to the advanced settings page.

In the 0.86.2 source code, lookup has two stages: first an exact match against 313 full names (such as gemini/gemini-3-pro-preview or openai/gpt-5.2); only if nothing matches does it move on to substring rules that catch, among others, sonnet-4-5, haiku-4-5, and gpt-4.1. To use an OpenAI-compatible gateway, the ID must be given as openai/<id>, and all 40 entries in the table with the openai/ prefix are models from OpenAI itself. So openai/claude-sonnet-4-6 matches nothing, and Aider uses whole-file rewriting, without a repo map: on each turn, the model returns every edited file in full. Output tokens are the expensive side of the bill.

Two lines fix this. Aider reads this file from your home directory, the git root, or the current directory, using the full name passed in --model as the key:

# name precisa ser idêntico à string passada em --model, com o prefixo
- name: openai/claude-sonnet-4-6
  edit_format: diff
  use_repo_map: true

For a one-off run, --edit-format diff is enough. By contrast, openai/claude-haiku-4-5 matches haiku-4-5 by substring and automatically enables diff format and the repo map—a naming coincidence that could change in any Aider version, so check rather than rely on it.

There is a second, quieter effect. Aider calls the API in three roles: the main model, an editor model used in --architect mode, and a weak model for commit messages and history summaries (see the options reference). With an unrecognized ID, the code uses the main model as both the weak and editor models, so all three roles are billed at the most expensive rate. With --weak-model, you can specify a cheaper ID (with the same openai/ prefix).

Finally, an unrecognized ID also hides the cost on screen: without input_cost_per_token, Aider shows the tokens without a cost line, and the warnings page confirms that it “assumes the model is free.” A metadata file fixes this:

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

Even with the metadata, Aider’s amount is its own calculation based on the reported tokens, with fixed cache multipliers for a single provider hard-coded in the code. Use it as an estimate and check your provider’s statement.

An estimate per session

The following is token arithmetic, not a measured task cost or an invoice ceiling. Suppose a session sends 200,000 uncached input tokens and receives 12,000 output tokens in diff format. In the whole-file column, suppose the same work produces 40,000 output tokens because the edited files are returned in full—an illustrative ratio, not a measurement. The rates are the current prices in the Kunavo catalog per million tokens.

With these assumptions, the edit format alone takes Claude Sonnet 4.6 from $0,546 to $0,840 for the same work: fix the setting before comparing providers. Commit message and summary traffic (suppose 30,000 input tokens and 2,000 output tokens) costs $0,084 if it stays on Claude Sonnet 4.6, versus $0,028 if you specify Claude Haiku 4.5 as the weak model.

Multiply by your number of sessions per day before treating this as a budget. Kunavo’s catalog amount is a billing floor, not a ceiling: when the upstream provider reports a cost, the charge is the greater of the catalog cost and the upstream cost times the applicable margin. Cache and external tools are outside this example.

“The cheapest API” is two questions

Aider publishes the only task-cost reference for this tool that can be independently checked: the polyglot benchmark runs the same 225 Exercism exercises in six languages with each model and records the cost in dollars.

Values read from the leaderboard data file. Two caveats matter more than the ranking: the latest entry is from October 3, 2025, so nothing released later was measured there; and the cost column gives no pricing schedule or date—it is what that run cost on that day, not a price you can buy at today. These models are cited as benchmark evidence; DeepSeek is not in the Kunavo catalog, and the inexpensive equivalent in the previous table is Claude Haiku 4.5.

The useful rule: o3-pro cost five times as much as gpt-5 (high) and scored lower, so a high rate does not buy completion; and a low rate does not guarantee the lowest total either, because a cheap model that needs three attempts can cost more than an expensive one that gets it right the first time. Choose the cheapest model that completes your work with an acceptable review effort, and confirm against the amount recorded in your account for that task. Aider’s /tokens measures, according to the command reference, the current conversation context, not the total already billed.

Which route to choose, and when

Two limits on the gateway route. A gateway without its own module connects through the compatibility route documented on the OpenAI-compatible APIs page: OPENAI_API_BASE, OPENAI_API_KEY, and the required openai/ prefix. There is no documented Anthropic-format base URL override, and the string ANTHROPIC_API_BASE does not appear in the 0.86.2 code. Also, do not count on a cache discount on this route: the cache page limits the feature to some provider families, requires the --cache-prompts option, and says cache statistics and costs are not shown with streaming. This page did not test whether cache survives the compatibility route.

Use Aider with Kunavo and pay with Pix

Kunavo publishes a setup guide for this client. It is a configuration reference, not a compatibility test: Aider was not run against Kunavo’s endpoint, and everything above was read from Aider’s published code and documentation. Keep a route that already works, run a bounded task, and then check the amount recorded in your account.

Payment is by prepaid top-up, with no monthly fee. The minimum top-up is $10 — the minimum needed to open a balance, not a per-task price or a subscription. Checkout is handled by Stripe: buyers in Brazil see Pix as an option, with the amount converted and displayed in reais before confirmation; cards, Apple Pay, Google Pay, and Link also work. The balance does not expire and failed requests are not charged. Details are in billing; when you are ready, create an account. The complete English version of this guide is Aider pricing.

Frequently asked questions

Aider repository, package, documentation, and leaderboard data verified on October 1, 2026; the behavior described was read in the published v0.86.2 code. Kunavo rates come from the live catalog, and every dollar example is illustrative token arithmetic, not a measured task cost.