cb1f1818fb
Real-world coding benchmark: Kimi K3 vs Fable 5 on 3D stadium challenge. - raw: raw/xpost/2026-07-20-thebuggeddev-kimi-vs-fable.md - wiki: kimi-k3.md +benchmark section, vibe-coding-vs-enterprise.md +xref, fable-5-anthropic.md +xref, moonshot-ai.md +Vibe Engineering section - index.md 80. Update, log entry
135 lines
6.9 KiB
Markdown
135 lines
6.9 KiB
Markdown
---
|
||
created: 2026-07-02
|
||
updated: 2026-07-05
|
||
sources: [xpost/2026-07-02_atomicchat-coding-benchmark-fable5-gpt55-opus48-glm52.md, wiki/tools/anthropic-claude.md, youtube/2026-07-05_everlast-ki-news-china-roboter-sonnet5-fable5.md]
|
||
tags: [concept, llm, fable5, anthropic, coding-model, frontier, premium-pricing, export-controls, cybersecurity, redeployment, figma-mcp, hixfield, remote-labor-index, verticalization]
|
||
---
|
||
|
||
# Fable 5 (Anthropic) — Frontier Coding Model, Premium Pricing
|
||
|
||
> **TL;DR:** Anthropic's Fable 5 is the public-facing version of Mythos 5 (cybersecurity-specialized). Redeployed 01.07.2026 with restrictions after June export-control crisis. Frontier coding quality (A+ in one-shot tests), but at **39× the cost of GLM 5.2** — the most expensive frontier model per token in our benchmark data.
|
||
|
||
## Overview
|
||
|
||
| Property | Detail |
|
||
|----------|--------|
|
||
| **Vendor** | Anthropic |
|
||
| **Status** | Redeployed (restricted) — back online 01.07.2026 |
|
||
| **Original launch** | 09.06.2026, blocked 12.06.2026 |
|
||
| **Relation to Mythos 5** | Public version of Mythos 5 (cybersecurity model) |
|
||
| **Pricing** | Premium — $3.12 for 62K tokens in atomic.chat benchmark |
|
||
| **Cost per 1K tokens** | $0.050 (highest in our dataset) |
|
||
| **Quality grade** | A+ (one-shot HTML5 Canvas physics, atomic.chat) |
|
||
|
||
## Export Control Crisis (June 2026)
|
||
|
||
Fable 5 was at the center of the June 2026 export control crisis:
|
||
|
||
1. **09.06.2026:** Fable 5 public launch
|
||
2. **12.06.2026:** Blocked by Trump administration after Amazon researchers jailbroke it via prompt injection
|
||
3. **01.07.2026:** Redeployed with "massive restrictions"
|
||
|
||
See [[../../tools/anthropic-claude.md]] for full timeline and [[../policy/ai-regulation-2026.md]] for regulatory context.
|
||
|
||
## Coding Benchmark Performance (atomic.chat, 01.07.2026)
|
||
|
||
| Metric | Value |
|
||
|--------|-------|
|
||
| Tokens generated | 62,158 |
|
||
| Cost | $3.12 |
|
||
| Grade | **A+** (best in test) |
|
||
| Scenes won | All three (train derailment, mid-air collision, monster truck) |
|
||
| vs GLM 5.2 | 39× more expensive, 2 quality grades higher (A+ vs B+) |
|
||
| vs GPT 5.5 | 2.7× more expensive, 1 quality grade higher (A+ vs A−) |
|
||
| vs Opus 4.8 | 5.6× more expensive (Opus grade unspecified) |
|
||
|
||
**Key finding:** Fable 5 produces the best raw output, but at a cost premium that is **hard to justify** for production use cases where B+ quality is sufficient. The 39× cost spread vs GLM 5.2 is the largest measured in our wiki.
|
||
|
||
> See [[coding-benchmark-price-performance.md]] for full benchmark analysis.
|
||
|
||
## Positioning
|
||
|
||
Fable 5 occupies the **premium frontier** position:
|
||
|
||
| Dimension | Fable 5 | Competition |
|
||
|-----------|---------|-------------|
|
||
| Quality | **A+** (best measured) | GPT 5.5 (A−), GLM 5.2 (B+) |
|
||
| Cost | **Highest** ($3.12/62K) | GLM 5.2 ($0.08/36K) |
|
||
| Access | Restricted (export controls) | GLM 5.2 open-weights (MIT) |
|
||
| Specialization | Cybersecurity (via Mythos 5 lineage) | General coding |
|
||
|
||
**The premium question:** Is A+ worth 39× the cost of B+? For most production use cases, no. For mission-critical, zero-failure-tolerance tasks — possibly. The answer depends on the cost of verification vs. the cost of generation.
|
||
|
||
## Update: Fable 5 Praxis-Test & Comeback-Details (05.07.2026)
|
||
|
||
**Everlast AI** (Leo Schmedding) liefert detaillierte Praxis-Einblicke in Fable 5 nach dem Comeback:
|
||
|
||
### Fable 5 Low vs Opus 4.8 Max
|
||
|
||
Fable 5 Low ist **günstiger, besser und schneller** als Opus 4.8 Max. Damit ist Fable 5 Low für die meisten Anwendungen das bevorzugte Standardmodell — zumindest bis zum 7. Juli 2026, in der es in normalen Plänen verfügbar ist.
|
||
|
||
### Figma MCP Praxis-Test (Marcel, Senior Developer)
|
||
|
||
**Task:** Komplettes Frontend-Design für eine native Tauri-App (Corporate LM) via Figma MCP.
|
||
|
||
| Metrik | Wert |
|
||
|--------|------|
|
||
| **Zeit** | 1,5 Stunden |
|
||
| **Effort** | Ultra Code (parallele Agenten in Figma) |
|
||
| **Basis** | Nur Code, keine Screenshots |
|
||
| **Output** | Cover, Foundations, Flows, App Shell, Empty/Befüllte States |
|
||
|
||
**Highlights:**
|
||
- Erkannte Tauri-Applikation (macOS + Windows) ohne explizite Nennung
|
||
- Erkannte Satoshi Font → Hinweis, Inter in Figma durch Satoshi zu ersetzen
|
||
- 1:1 Übernahme der Kernscreens aus Webapp-Code
|
||
- Früher: Designteam Wochen/Monate → Fable 5: 1,5 Stunden
|
||
|
||
### Fallback zu Opus 4.8 deaktivierbar
|
||
|
||
Standardmäßig wird bei "normalen" Coding-Aufgaben zu Opus 4.8 geforwarded. **Deaktivierbar:** Claude Settings → Fähigkeiten → "Modell wechseln, wenn eine Nachricht markiert wird" → ausschalten. Chat wird dann pausiert statt weitergeleitet — verhindert Token-Verschwendung.
|
||
|
||
### Hixfield Integration
|
||
|
||
Fable 5 kann über **Hixfield MCP** für Erklärvideos genutzt werden (Hixfield Explainer).
|
||
|
||
### Community-Feedback (durchwachsen)
|
||
|
||
- Einige: Fable 5 vor dem US-Ban besser als nach Wiederkunft
|
||
- AI Arena (seriös): Fable 5 nach Re-Release in einigen Bereichen **besser** (Dokumente, Creative Writing)
|
||
- Für den Massenmarkt: Keine großen Sprünge mehr spürbar
|
||
- Prognose: Anthropic könnte $500/$1.000 Plan einführen
|
||
|
||
### Remote Labor Index — Fable 5 CAD-Stärke
|
||
|
||
Der Remote Labor Index (240 Freelance-Projekte: Grafikdesign, Architektur, CAD, Video, Audio, Data Analysis, Webdevelopment) zeigt: **CAD ist ein riesen Use Case für Fable 5**. Der "Remote Turing Test" wird 2026 als bestanden prognostiziert.
|
||
|
||
### Claude Science — Medikamentenentwicklung
|
||
|
||
Anthropic steigt offiziell in **Drug Development** ein (Claude Science). Nachdem Mathematik, Physik und Coding "gelöst" sind, kommt Biologie/Medikamentenentwicklung als nächste Disziplin.
|
||
|
||
### Vertikalisierung — Claude Produkt-Linie
|
||
|
||
**Bestehend:** Claude Code, Claude Cowork, Claude Design, Claude Finance, Claude Science
|
||
**Ausstehend:** Claude HR, Claude Analytics, Claude Marketing, Claude Sales, Claude Legal, Claude Logistics, Claude CAD, Claude R&D, Claude Accounting
|
||
|
||
**Trend:** Vertikalisierung von KI-Anwendungen.
|
||
|
||
## Cross-References
|
||
|
||
- [[../../tools/anthropic-claude.md]] — Anthropic Claude family overview, export control timeline
|
||
- [[coding-benchmark-price-performance.md]] — Full atomic.chat benchmark analysis
|
||
- [[../policy/ai-regulation-2026.md]] — Export control regulatory context
|
||
- [[../../people/dario-amodei.md]] — Anthropic CEO, regulatory stance
|
||
- [[glm-5.2-zai-coding-model.md]] — 39× cheaper competitor
|
||
- [[chinese-model-cost-routing.md]] — Cost-routing thesis
|
||
- [[sonnet-5-anthropic.md]] — Sonnet 5 (teurer als Fable 5 im Cost-to-Run — Ironie)
|
||
- [[../hardware/ubtech-u1-humanoid-roboter.md]] — Cloud-KI als Roboter-Backend
|
||
- [[vibe-coding-vs-enterprise.md]] — Agentic Coding / Vibe Coding
|
||
- [[kimi-k3.md]] — Kimi K3 vs. Fable 5 real-world benchmark (20.07.2026): Fable 5 faster but less thorough; Kimi K3 did E2E testing, responsive validation, clean React architecture
|
||
|
||
## External Sources
|
||
|
||
- [atomic.chat benchmark X-Post](https://x.com/atomic_chat_hq/status/2072446067962978411)
|
||
- [Anthropic](https://www.anthropic.com)
|
||
- [Everlast AI — KI-News 05.07.2026](https://www.youtube.com/watch?v=-JCCcR9qtYQ)
|