Aider works with any OpenAI-compatible model — which means you have hundreds of options. But which model actually produces the best results? We ranked the top 8 based on Aider’s own polyglot benchmark (82.3%), real-world coding quality, cost per session, and developer experience.
The rankings
#1: DeepSeek V4-Pro — Best overall value
aider --model openrouter/deepseek/deepseek-v4-pro| Quality | Cost | Speed |
|---|---|---|
| 80.6% SWE-bench Verified | $0.435/$0.87 per M | ~100 t/s |
DeepSeek V4-Pro gives you the highest coding benchmark scores at the lowest price. A typical Aider session costs $0.05-0.15. Monthly heavy use: $5-15. The permanent discount makes this the default recommendation.
Best for: Daily coding, refactoring, bug fixing. The model you should start with.
#2: Claude Opus 4.8 — Best absolute quality
aider --model claude-opus-4-8| Quality | Cost | Speed |
|---|---|---|
| 69.2% SWE-bench Pro (harder benchmark) | $5/$25 per M | ~80 t/s |
Claude Opus 4.8 produces the most reliable code with 4× fewer unflagged errors. Expensive but worth it for production-critical changes and complex multi-file refactoring.
Best for: Complex architecture changes, production code, when you can’t afford bugs.
#3: MiMo V2.5 Pro — Best for long sessions
aider --model openrouter/xiaomi/mimo-v2.5-pro| Quality | Cost | Speed |
|---|---|---|
| 79.2% SWE-bench, 97.2% tool accuracy | $0.435/$0.87 per M | ~80 t/s |
MiMo V2.5 Pro uses 40-60% fewer tokens than competitors. Your Aider sessions cost less AND go further before hitting context limits. Designed for long editing sessions with many back-and-forth exchanges.
Best for: Long refactoring sessions, iterative development, budget optimization.
#4: Qwen 3.6 27B (local via Ollama) — Best free option
aider --model ollama/qwen3.6:27b| Quality | Cost | Speed |
|---|---|---|
| Strong (competitive with larger models) | FREE (local) | 25-40 t/s |
Qwen 3.6 27B via Ollama gives you solid coding assistance at zero cost. Needs 16GB+ RAM (Q4 quantization). NVIDIA optimized this model for RTX Spark with 2× throughput.
Best for: Offline development, privacy, zero-budget coding.
#5: Qwen 3.7 Max — Best reasoning
aider --model openrouter/qwen/qwen3.7-max| Quality | Cost | Speed |
|---|---|---|
| 92.4% GPQA, ~58% SWE-Pro | $2.50/$7.50 per M | Standard |
Qwen 3.7 Max thinks deepest about complex problems. Use it for architecture decisions, algorithm design, and tasks requiring multi-step reasoning that cheaper models struggle with.
Best for: Complex algorithms, system design, mathematical computing.
#6: MiniMax M3 — Best multimodal with Aider
aider --model openrouter/minimax/minimax-m3| Quality | Cost | Speed |
|---|---|---|
| 59% SWE-Pro, 83.5% BrowseComp | $0.60/$2.40 per M | Fast (MSA) |
MiniMax M3 adds vision — useful when you want to show Aider a screenshot of a bug or a mockup to implement. 1M context + MSA speed for large codebases.
Best for: Visual context (screenshots, diagrams), large codebase analysis.
#7: DeepSeek V4 Flash — Absolute cheapest
aider --model openrouter/deepseek/deepseek-v4-flash| Quality | Cost | Speed |
|---|---|---|
| Good (distilled from V4-Pro) | $0.07/$0.28 per M | ~150 t/s |
When you want Aider for simple tasks (rename variables, add comments, small fixes) and don’t want to spend anything. A full day of light use costs pennies.
Best for: Simple edits, learning Aider, high-volume batch work.
#8: Llama 4 Scout (local) — Best large local model
aider --model ollama/llama4-scout| Quality | Cost | Speed |
|---|---|---|
| Good (109B knowledge, 17B active) | FREE (local) | 10-15 t/s |
Llama 4 Scout packs 109B parameters of knowledge into a model that runs at 17B speed. Needs 60GB+ RAM. Great for local use on high-memory machines (RTX Spark, Mac Studio).
Best for: Large local model with broad knowledge, offline use.
Quick setup
All models work with Aider via OpenRouter (one API key):
# Install Aider
pip install aider-chat
# Set OpenRouter key
export OPENROUTER_API_KEY="your-key"
# Use any model
aider --model openrouter/deepseek/deepseek-v4-proFor direct API access (skip OpenRouter markup), see:
Cost comparison (1hr coding session)
| Model | ~Cost/hour | Monthly (4hr/day) |
|---|---|---|
| DeepSeek V4 Flash | $0.02 | $2 |
| DeepSeek V4-Pro | $0.08 | $10 |
| MiMo V2.5 Pro | $0.06 | $8 |
| MiniMax M3 | $0.15 | $18 |
| Qwen 3.7 Max | $0.50 | $60 |
| Claude Opus 4.8 | $2.25 | $270 |
| Ollama (local) | $0.00 | $0 |
FAQ
Which model should I start with?
DeepSeek V4-Pro via OpenRouter. Best quality-to-cost ratio. Use aider --model openrouter/deepseek/deepseek-v4-pro. Upgrade to Opus only if you hit quality limits.
Can I switch models mid-session?
Yes. Use /model command in Aider to switch without restarting. Useful for escalating: do routine work with DeepSeek, switch to Opus for the hard part.
Do local models work well with Aider?
Qwen 3.6 27B is excellent for most tasks. Models smaller than 14B tend to struggle with multi-file edits. See best Ollama models for coding.
What about Aider’s own benchmarks?
Aider’s polyglot benchmark (82.3%) was tested with Claude. Other models score differently. DeepSeek and MiMo produce comparable results in practice. The benchmark mainly validates Aider’s editing format compatibility with each model.
Claude Code or Aider — which is better?
Claude Code has dynamic workflows and subagents but locks you to Claude ($5/$25). Aider is free and works with any model. See our full comparison guide. Most developers use Aider for budget work and Claude Code for complex tasks.