Claude Haiku 5.5 vs GPT-6 Luna: same sticker price, but a 5x cliff past 100K tokens
大模型之路 · wechat · 2026-10-09
Anthropic priced Claude Haiku 5.5 identically to GPT-6 Luna ($0.10/$0.50 per M tokens), but the article breaks down the real engineering math: pricing jumps 5x past 100K prompt tokens (Luna tolerates up to 272K), the new tokenizer adds 30% tokens for Chinese, and effort tiers act as cost valves. Sub-agent routing (Opus 5.5 + 10 Haiku 5.5 workers) cut demo cost by 70% at 3.7x speed, and cache reads at $0.01/M are an overlooked saving. Includes migration pitfalls: default thinking, 400 errors on non-default sampling params.
Related event: Claude Haiku 5.5 ranks 2nd in intelligence but real costs spiral(16 posts)→
More from coding & agent
- 10 skills to fix the cookie-cutter look of AI-generated websites, drop into Claude Code or Cursor — sven_ai · 2026-11-22
- A catalog of 2,256 real agent use cases with copyable prompts, from bill-negotiating bots to 11-agent trading teams — sven_ai · 2026-11-22
- Disney Parks MCP Server offers structured queries for attractions, dining and Lightning Lane status — modelcontextprotocol · 2026-11-22
- Trade Intel MCP chains Comtrade, Census, and Treasury data for agents — modelcontextprotocol · 2026-11-22
- DeepSeek Harness v0.2.3-alpha.1 ships 14 changes: plugin market, PR review, multi-agent UI — teortaxesTex · 2026-11-22
- 2-day Gemini stress test: $4.99 Google AI Pro nearly maxed, orchestrator UI revealed — lipeng0820 · 2026-11-22