Aider: 실제 토큰 배수 측정 및 BYOK 비용 절감 가이드
Token-frugal open-source CLI using Tree-sitter AST Repo-Map, Architect/Editor split, and clean Git auto-commits.
Context & Token Overhead Mechanics
In a 15-turn refactor with 5 tool calls, AST Repo-Map + search/replace diffs keep total multiplier at just 1.2x–1.8x (the lowest in this benchmark).
Baseline instructions & tool schema weight per turn
Tree-sitter AST Repo-Map (1k–2k token budget) + search/replace diff edits + explicit /drop context hygiene
Pass `--cache-prompts` (plus `--cache-keepalive-pings` for Anthropic) to cache Repo-Map and added read-only files.
Model Compatibility Matrix & 15-Turn Cost
Estimated on 30k raw input + 4.5k output scaled by 1.4x multiplier at 70% cache hit rate
| Model | Tool Calling | Prompt Caching | Vision | Streaming | Est. 15-Turn Cost |
|---|---|---|---|---|---|
| DeepSeek-V3 (deepseek-chat)Budget Workhorse | Diff/Edit Block | Full | N/A | Stable | $0.011 |
| DeepSeek-R1 (deepseek-reasoner)Architect Mode | Architect Only | Full | N/A | Stable | $0.023 |
| Anthropic: Claude Sonnet 5.5Primary Editor | Native | Full | Supported | Stable | $0.082 |
| Google: Gemini 4 NeonHigh-Speed Large Repo | Native | Full | Supported | Stable | $0.123 |
Production BYOK & DeepSeek / OpenRouter Config
Built on LiteLLM: supports 100+ providers, custom OpenAI-compatible endpoints, local Ollama, and dual-model Architect/Editor routing out of the box.
model: deepseek/deepseek-chat
architect: true
editor-model: deepseek/deepseek-chat
cache-prompts: true
map-tokens: 1024
auto-commits: true