Windsurf:実測トークン倍率&BYOK コスト最適化ガイド
Flow-driven AI IDE powered by Cascade engine with selective AST chunking and strict step-by-step tool control.
Context & Token Overhead Mechanics
In a 15-turn refactor with 5 tool calls, Cascade's selective chunk retrieval keeps overhead at 1.9x–2.8x—roughly 15%–20% leaner than unpruned multi-file IDE agents.
Baseline instructions & tool schema weight per turn
Cascade flow-aware selective chunking + automatic terminal output truncation
Cascade orders repository context more deterministically than naive RAG, yielding solid cache reuse on Anthropic and Gemini models.
Model Compatibility Matrix & 15-Turn Cost
Estimated on 30k raw input + 4.5k output scaled by 2.3x multiplier at 70% cache hit rate
| Model | Tool Calling | Prompt Caching | Vision | Streaming | Est. 15-Turn Cost |
|---|---|---|---|---|---|
| Anthropic: Claude Sonnet 5.5Primary | Native | Full | Supported | Stable | $0.115 |
| Google: Gemini 4 NeonHigh-Speed Agent | Native | Full | Supported | Stable | $0.172 |
| OpenAI: GPT-6.2 SolGeneral Coding | Native | Full | Supported | Stable | $0.115 |
| DeepSeek-V3 (deepseek-chat)Available via Windsurf hosted tier / custom config | Native | Provider-Dependent | N/A | Stable | $0.016 |
Production BYOK & DeepSeek / OpenRouter Config
Supports bringing personal Anthropic/OpenAI/Google keys for select models, while arbitrary third-party Base URL overrides are restricted compared to open extensions.
{
"googleApiKey": "${GEMINI_API_KEY}",
"anthropicApiKey": "${ANTHROPIC_API_KEY}",
"defaultCascadeModel": "gemini-4-neon"
}