Task 090: Cost Analysis — Non-Max-Subscription Estimates

Status: DONE

Motivation

Human directive: コスト分析においては、実際はmax subscriptionでもそうでなかった場合のコスト見積もりもすること

Completed

Added "プラン別コスト比較" section to ai-costs.json with:

Plans compared

1. **Max Plan ($100-200/mo)**: ~$7 per VM boot (Haiku subagent API only)
2. API-only: ~$196 per VM boot (all tokens at API rates)
3. Batch API: ~$195 (50% I/O discount, non-interactive only)
4. Pro Plan ($20/mo): Rate-limited, not practical for agent loop

Key findings

• Cache is critical: 97.3% cache hit rate reduces Opus effective rate to $0.50/MTok (10% of base)
• Max Plan ROI: $100-200/mo subscription vs ~$196 API cost — pays for itself in one VM boot
• Per-session cost: $3.39 (API) vs $0.12 (Max) — 28× difference
• Opus dominates cost: Main session Opus tokens (260M) account for 89% of API cost

Pricing sources

• Official API pricing from platform.claude.com/docs/en/about-claude/pricing (Feb 2026)
• Opus 4.6: $5/$25/MTok (input/output), $0.50 cache read, $6.25 cache write
• Haiku 4.5: $1/$5/MTok, $0.10 cache read, $1.25 cache write