Codex 校核（2026-10-10）：以下更正适用于本页；原采集内容保留为历史快照。
- 25/MTok 对比的是每百万输出 token 单价，不能当作整任务价格；输入、缓存、推理输出量与服务商计费都会影响总成本。数字保留为原文历史报价，本文未据此证明所有任务更省钱。Fireworks 官方 GLM 5.2 价表、Anthropic 对 Opus 4.8 同价的官方说明
- Wafer 的 2.75× 指 MI355X 相对 B300 的每 GPU 价格差，不是通用“每 token 便宜 2.75×”。其 GLM 5.2 推理实验报告超过 2× 的成本优势，聚合测试采用 20k 输入／1k 输出、60% 缓存命中等指定条件；不能推广为所有 AMD 对 Blackwell 的推理成本。原始部署实验
Docx-CLI + GLM 5.2 — Knowledge Digest
Finding 1: Docx-CLI — AI agents reading/editing Word docs
Source: HN Show HN 48821500 | Repo: kklimuk/docx-cli | ⭐22
A .docx CLI built for AI agents. Agents edit Word docs via stable locators (p3:5-20) — tracked changes, comments, redlines survive. Mutates OOXML in-place rather than re-emitting from a lossy LLM.
Benchmark highlights (6 real doc tasks):
- Haiku + docx-cli: 4.3/6 vs default skill 0.7/6 (6× improvement)
- Sonnet + docx-cli: 6/6 vs default skill 4/6
- 0 broken documents vs ~1/run for default skill
- 2.2-2.6× fewer tokens, 1.7-2.0× faster
- Ships as Agent Skill format — compatible with Claude Code, Codex, Pi
Relevance to Hermes: Direct integration possible — Hermes currently has limited .docx handling. The CLI + Skill format could be adopted for document processing pipelines.
Finding 2: GLM 5.2 and the coming AI margin collapse
Source: HN 48809877 (653pts) | Article: martinalderson.com
GLM 5.2 (25/MTok) for agentic coding. Switching costs are near-zero — Z.ai and Fireworks provide both OpenAI-compatible and Anthropic-compatible endpoints.
Key analysis:
- 15-20% of Opus/GPT 5.5 price, similar quality for agentic tasks
- No native vision or web search (gap being filled by MCP servers)
- AMD inference 2.75× cheaper per token than Nvidia Blackwell
- 90% gross margins on inference API unsustainable
- Part 2 promises industry impact analysis
Relevance to Hermes: Multi-model provider strategy validated. GLM 5.2 viable as cheap backend for non-interactive agentic tasks (PR review, batch processing).