aee635ef50
New raw file: raw/blog/2026-07-02_whytryai-language-blind-spot.md New wiki page: wiki/concepts/llm/ai-search-language-blind-spot.md - Language Lens Bias concept + Local Sources Prompt hack - Denmark case study (English vs Danish search, ~0% topic overlap) - Global Pulse Claude Code Skill for multi-country comparisons - Routing parallel to chinese-model-cost-routing (model routing vs source routing) - Lokal vs. Global cross-section thesis expanded (5th dimension: source routing) Cross-refs added to: mlx-moe-local-ai-optimization, local-llm-laptop-guide, chinese-model-cost-routing Index updated (57. Update), log entry added.
164 lines
No EOL
11 KiB
Markdown
164 lines
No EOL
11 KiB
Markdown
---
|
||
created: 2026-06-29
|
||
updated: 2026-07-02
|
||
sources:
|
||
- xpost/2026-06-29_deronin-chinese-ai-stack-cost-savings.md
|
||
- xpost/2026-06-28_milesdeutscher-token-engineering-barbell.md
|
||
- xpost/2026-07-02_atomicchat-coding-benchmark-fable5-gpt55-opus48-glm52.md
|
||
tags: [concept, llm, chinese-models, cost-routing, barbell, model-swap, kimi, qwen, glm, mimo, wan, kling, cost-reduction, factory-for-gods, practical-report, ct-validation]
|
||
---
|
||
sources_extra: [youtube/2026-07-02_ct4004-anthropic-ceo-panik-open-weights-poisonai.md]
|
||
|
||
# Chinese Model Cost Routing — The 87% Cost-Cut Playbook
|
||
|
||
> **TL;DR:** DeRonin's 30-day field report swaps six Western frontier models for Chinese alternatives — 87% cost reduction, 4% quality drop, revenue unchanged. This is the empirical validation of the Barbell Model Routing strategy and TheProphet's "Factory for Gods" thesis.
|
||
|
||
## Sources
|
||
|
||
| Source | Author | Date | Angle |
|
||
|--------|--------|------|-------|
|
||
| [X-Post: "My entire AI stack is now Chinese"](https://x.com/DeRonin_/status/2071561335234531578) | DeRonin (@DeRonin_) | 2026-06-29 | Practical field report |
|
||
| Miles Deutscher "Token Engineering" | @Milesdeutscher | 2026-06-28 | Barbell strategy formalization |
|
||
| [TheProphet "Factory for Gods"](https://x.com/_The_Prophet__/status/2067390526157185188) | @TheProphet | 2026-06-17 | Macro-thesis: China industrializes intelligence |
|
||
|
||
## The Swap Matrix
|
||
|
||
DeRonin replaced every layer of his AI stack with Chinese models:
|
||
|
||
| Task | Western (before) | Chinese (after) | Quality Gap | Cost Factor | Notes |
|
||
|------|------------------|-----------------|-------------|-------------|-------|
|
||
| Reasoning / Backend Brain | Claude Opus 4.8 | **Kimi K2.7** (Moonshot) | ~8% | ~11× cheaper | Largest absolute saving — Opus is premium-priced |
|
||
| Code Generation | GPT-5.5 | **Qwen 3.7 Max** (Alibaba) | ~18% | ~7× cheaper | Biggest quality gap, but still "good enough" for production |
|
||
| Agent Loops + Tool Calling | Claude Sonnet 4.7 | **GLM 5.2** (Z.ai) | ~3% | ~5× cheaper (input) | Smallest gap — GLM 5.2 is near-parity for agentic tasks |
|
||
| Cheap Volume / Bulk | GPT-5.5 mini | **MiMo V2.5** (Xiaomi) | ~6% | ~12× cheaper | Best cost ratio in the stack |
|
||
| Image Generation | GPT-Image-2 | **Wan 2.5** | ~5% | ~8× cheaper | — |
|
||
| Video Generation | Sora 2 | **Kling 3.0** (Kuaishou) | ~equal | ~6× cheaper | Roughly equal quality at a fraction of cost |
|
||
|
||
### 30-Day Outcome
|
||
|
||
| Metric | Before | After | Delta |
|
||
|--------|--------|-------|-------|
|
||
| Operating costs | baseline | 13% of baseline | **−87%** |
|
||
| Output quality (avg) | baseline | 96% of baseline | **−4%** |
|
||
| Revenue | baseline | unchanged | **0%** |
|
||
|
||
**Key insight:** The 4% quality drop is across all tasks averaged. For agent loops (GLM 5.2) and video (Kling 3.0), the gap is ≤3% — effectively parity. The biggest gap (Qwen 3.7 Max for code gen at 18%) is offset by the fact that code generation volume is high and verification is cheap.
|
||
|
||
## Connection to Barbell Model Routing
|
||
|
||
DeRonin's stack is a **natural implementation of the Barbell Strategy** (Miles Deutscher, 2026-06-28):
|
||
|
||
| Barbell Phase | DeRonin's Implementation | Model |
|
||
|---------------|--------------------------|-------|
|
||
| **First 10%** (Planning / expensive) | Reasoning tasks where 8% quality gap matters | Kimi K2.7 (cheapest "brain") |
|
||
| **Middle 80%** (Execution / cheap) | Agent loops, bulk processing, code gen | GLM 5.2, MiMo V2.5, Qwen 3.7 Max |
|
||
| **Last 10%** (Verification / expensive) | Implicit — revenue unchanged means verification held | (not detailed in post) |
|
||
|
||
### Our Setup vs. DeRonin's
|
||
|
||
| Task | DeRonin | Hector (us) | Match? |
|
||
|------|---------|-------------|--------|
|
||
| Primary agent brain | Kimi K2.7 | GLM 5.2 (cloud) | Different — we route reasoning to GLM 5.2 |
|
||
| Code generation | Qwen 3.7 Max | Kimi K2.7 Code (subagent) | Different — we use Kimi for code |
|
||
| Agent loops + tools | GLM 5.2 | GLM 5.2 (native in OpenClaw v2026.6.8) | ✅ Same |
|
||
| Bulk processing | MiMo V2.5 | MiMo in fallback chain | ✅ Same family |
|
||
| Routing logic | (promised article) | `skills/openclaw-model-router` 5-tier | We have formalized routing |
|
||
|
||
**Where we differ:** DeRonin uses Kimi K2.7 for reasoning and GLM 5.2 for agent loops. We use GLM 5.2 as the primary for everything including reasoning, with Kimi K2.7 Code as a subagent option. His split is more granular — reasoning gets a dedicated "brain" model, agent loops get the cheaper agentic model. Our 5-tier router achieves similar granularity via task-complexity classification rather than task-type classification.
|
||
|
||
## Connection to "Factory for Gods" Thesis
|
||
|
||
This is the **micro-level proof** of TheProphet's macro-thesis (see [[ai-intelligence-commoditization-thesis.md]]):
|
||
|
||
> *"America has the frontier gods. China is building the factory for gods."*
|
||
|
||
DeRonin's 30-day report demonstrates empirically what TheProphet argued theoretically:
|
||
|
||
1. **Near-frontier is good enough** — A 4% average quality drop is invisible to end-users and revenue-neutral
|
||
2. **Cost asymmetry is extreme** — 5–12× cheaper per task category, compounding to 87% total
|
||
3. **The swap is not hypothetical** — It's a production stack that ran for 30 days with unchanged revenue
|
||
4. **Every layer is covered** — Not just text generation, but code, agents, images, and video
|
||
|
||
### The Quality-Cost Curve in Practice
|
||
|
||
TheProphet's thesis predicts that "a 10% intelligence gap is economically irrelevant when the second-best model is radically cheaper." DeRonin's data refines this:
|
||
|
||
- **Agent loops:** 3% gap at 5× cheaper → **no-brainer swap**
|
||
- **Reasoning:** 8% gap at 11× cheaper → **clearly worth it**
|
||
- **Code gen:** 18% gap at 7× cheaper → **the interesting edge case** — still profitable because code verification is cheaper than code generation
|
||
- **Video:** ~0% gap at 6× cheaper → **pure arbitrage**
|
||
|
||
The threshold is not "10% gap" universally — it's task-dependent. Where verification is cheap (code), larger gaps are tolerable. Where it's expensive (reasoning), smaller gaps matter more.
|
||
|
||
## Chinese Model Landscape (Referenced)
|
||
|
||
| Model | Vendor | Category | In Our Stack? |
|
||
|-------|--------|----------|---------------|
|
||
| Kimi K2.7 | Moonshot AI | Reasoning / coding | ✅ Subagent option |
|
||
| Qwen 3.7 Max | Alibaba | Code generation | Qwen3 Coder in fallback |
|
||
| GLM 5.2 | Z.ai | Agent loops / tool calling | ✅ Primary model |
|
||
| MiMo V2.5 | Xiaomi | Bulk / cheap volume | ✅ Fallback chain |
|
||
| Wan 2.5 | Alibaba (Wan) | Image generation | ❌ Not integrated |
|
||
| Kling 3.0 | Kuaishou | Video generation | ❌ Not integrated |
|
||
|
||
See [[llm-model-catalog.md]] for full model inventory and [[glm-5.2-zai-coding-model.md]] for our GLM 5.2 deep-dive.
|
||
|
||
## Implications for Our Architecture
|
||
|
||
1. **GLM 5.2 as primary is validated** — DeRonin independently chose GLM 5.2 for agent loops, the task category with the smallest quality gap (3%). Our default model choice is optimal for the highest-volume task type.
|
||
|
||
2. **Kimi K2.7 for reasoning is worth considering** — DeRonin splits reasoning (Kimi) from agent loops (GLM). Our current setup routes both to GLM 5.2. A potential optimization: route complex reasoning tasks (Tier 4 in our 5-tier router) to Kimi K2.7 instead of GLM 5.2.
|
||
|
||
3. **MiMo for bulk is confirmed** — We already have MiMo in the fallback chain. DeRonin's 12× cost factor for bulk processing validates Tier 0/1 routing.
|
||
|
||
4. **Image/Video gap to explore** — We don't currently use Wan 2.5 or Kling 3.0. If image/video generation becomes a regular need, these are the cost-optimal choices.
|
||
|
||
5. **Full article pending** — DeRonin promised a detailed article with routing logic for 2026-06-30. This page should be updated when it drops.
|
||
|
||
## Cross-References
|
||
|
||
- [[ai-intelligence-commoditization-thesis.md]] — TheProphet's "Factory for Gods" macro-thesis
|
||
- [[ai-value-migration-orchestration.md]] — Aravind's investor angle on value migration
|
||
- [[flat-curve-society.md]] — Yegge's "route to the dumbest model that can handle it"
|
||
- [[glm-5.2-zai-coding-model.md]] — GLM 5.2 deep-dive (our primary = DeRonin's agent loop choice)
|
||
- [[real-world-coding-showdown.md]] — Kimi K2.7 vs GLM 5.2 head-to-head
|
||
- [[llm-model-catalog.md]] — Full model inventory
|
||
- [[../../architecture/model-routing.md]] — Our routing architecture
|
||
- [[../policy/ai-as-geopolitical-weapon.md]] — Miles Deutscher's weaponization thesis
|
||
- `skills/openclaw-model-router/SKILL.md` — 5-tier routing skill (external to wiki)
|
||
- [[ai-search-language-blind-spot.md]] — Routing-Parallele: Quellen-Routing (lokale Sprache) folgt derselben Logik wie Modell-Routing (bewusstes Routing schaltet verdeckte Qualität frei)
|
||
|
||
|
||
## c't 4004 #30 Validation (02.07.2026)
|
||
|
||
Die c't-Folge "Anthropic-CEO in Panik" (c't 4004 #30, 02.07.2026) liefert eine **unabhängige journalistische Bestätigung** der chinesischen Model-Cost-Routing-These:
|
||
|
||
- **Große Unternehmen wechseln** bereits weg von US-Cloud-Anbietern zu selbst gehosteten chinesischen Open-Weight-Modellen
|
||
- **GLM-5.2 explizit genannt** als Beispiel für Modelle, die den Rückstand zu US-Frontier bei dramatisch geringeren Kosten schließen
|
||
- **US-Sanktionspolitik als kurzsichtig kritisiert** — statt China aufzuhalten, beschleunigt sie unabhängige KI-Kapazitäten im Rest der Welt
|
||
- **Für Europe:** Chance zur Emanzipation von US-Hyperscalern
|
||
|
||
Diese journalistische Validierung aus dem Heise-Ökosystem (c't Magazin) ist signifikant, weil sie unabhängig von der X/Twitter-Community kommt und ein Mainstream-Tech-Publikum erreicht.
|
||
|
||
Quelle: [YouTube: c't 4004 #30](https://www.youtube.com/watch?v=Q6QkzJwcrEc) — Kapitel 16:55 "Chinas Open Weights killen die KI-Blase"
|
||
|
||
## atomic.chat Benchmark — 39× Cost Spread (01.07.2026)
|
||
|
||
Ein weiterer Datenpunkt, der die Cost-Asymmetry auf die Spitze treibt: [atomic.chat](https://x.com/atomic_chat_hq/status/2072446067962978411) testete 4 Modelle im One-Shot HTML5 Canvas Physics Benchmark.
|
||
|
||
| Model | Cost | vs GLM 5.2 |
|
||
|-------|------|------------|
|
||
| Fable 5 | $3.12 | **39× teurer** |
|
||
| GPT 5.5 | $1.14 | 14.25× teurer |
|
||
| Opus 4.8 | $0.56 | 7× teurer |
|
||
| **GLM 5.2** | **$0.08** | **1×** |
|
||
|
||
Die 39×-Asymmetry ist eine Größenordnung über DeRonin's pro-Task-Werten (5-12×). Grund: Fable 5 ist das teuerste Frontier-Modell (Mythos-5-Lineage Premium-Pricing), GLM 5.2 das günstigste — der Spread ist maximal. GLM 5.2 erreichte B+ Qualität (vs Fable 5 A+), blieb also competitive.
|
||
|
||
Siehe [[coding-benchmark-price-performance.md]] für die vollständige Analyse.
|
||
|
||
## External Sources
|
||
|
||
- [DeRonin X-Post (Original)](https://x.com/DeRonin_/status/2071561335234531578)
|
||
- [Miles Deutscher "Token Engineering" (2026-06-28)](https://x.com/Milesdeutscher)
|
||
- [TheProphet "Factory for Gods" (2026-06-17)](https://x.com/_The_Prophet__/status/2067390526157185188) |