Aider is free, so its whole cost is the model tokens your engineers spend. Assign the main, editor, and weak models in a shared config, approve the providers code may go to, and put a budget alert on every provider account.
When you need this
- Engineers each use a different model and provider.
- Finance asks what Aider costs per engineer.
- Code must stay on approved providers or on the machine.
- You are deciding whether cheaper or local models are good enough.
Evaluation criteria
Assign the model roles
Which model plans, which edits, and which writes commit messages?
Look forA main model, an editor model for architect mode, and a cheaper weak model for commit messages and history summaries, set in .aider.conf.yml. Aider picks the best edit format for known models, and less familiar models default to the whole-file format, which costs more tokens.
Approve the providers
Where may code be sent?
Look forA short list from what Aider supports: Anthropic, OpenAI, Gemini, DeepSeek, OpenRouter, Bedrock, Vertex AI, Azure, OpenAI-compatible gateways, a GitHub Copilot token, or local models through Ollama or LM Studio. Keys go in environment variables or a .env file, never a committed config.
Test local and cheaper models
Do budget models do your work?
Look forYour own trial on your own code. Some engineers report useful results from local Qwen and DeepSeek models in architect mode, and others found that free and small local models fail to produce edits. Aider's Ollama instructions raise the context window to 8,192 tokens, because a small window leaves no room for the repo map.
Make spend visible
Can an engineer see what a session costs while it runs?
Look forAider's per-message and per-session token and cost report, and /tokens before a large request. Budget alerts belong on each provider account or a gateway, since Aider has no spend limit of its own.
Budget from real usage
What does heavy use cost on your models?
Look forNumbers from your pilot. One engineer reports about 3 US cents a prompt on Gemini with context held under 25,000 tokens, and agentic use on frontier models costs far more. The leaderboard's benchmark costs show the spread between models on the same 225 tasks.
Compare against flat plans
At your usage, is per-token spend cheaper than a subscription?
Look forCost per accepted change on both. Four accounts prefer paying per token, and three prefer a flat subscription after API spend ran high or proved hard to predict.
Questions to ask vendors
- Which providers may engineers configure, and who audits the keys on each machine?
- Does our gateway need custom headers, which Aider sets per model in a model settings file?
- Which models has Aider's main branch added since the 0.86.2 release, and do we need them?
- How do we revoke a Copilot token or provider key when someone leaves?
Red flags
- Personal API keys with no budget alert on the provider account.
- A team standard set from 2025 leaderboard scores without testing current models.
- Local models run at a small context window with the full repo map.
- Spend reported only as a monthly total per provider.
Products in this piece
Sources
- Official documentation
- 8
- Public source
- 6
- Aider docs: Connecting to LLMs
- Aider docs: API Keys
- Aider docs: Ollama
- Aider docs: GitHub Copilot
- Aider docs: Reasoning models
- Aider LLM Leaderboards: Benchmark notes
- Aider docs: Options reference
- Aider LLM Leaderboards (last updated Nov 20, 2025)
- Hacker News, BeetleB: about 3 cents a prompt on Gemini under 25k tokens (Jun 12, 2025)
- Hacker News, gwd: cost per query and per session as a daily driver (Feb 24, 2025)
- Hacker News, regularfry: local architect and editor models on Ollama (Mar 10, 2025)
- Hacker News, lreeves: gpt-oss-120b locally against GPT-5 and Claude (Oct 31, 2025)
- Musa Nayyer, DEV Community: I Tried Building a Complex Security Tool with a 1.5B Local Model (May 27, 2026)
- GitHub issue #1977: a company proxy that needs headers on every model (Oct 8, 2024; not a Signals source)