Fable 5 is back… here is my plan

Tech/Product Analysis Summary (Auto-subs; Key Concepts Only)

1) Fable 5 “returned” — why it’s better for coding (vs other models)

• The main claim is that Fable 5 is the “best model to work with” and represents a step-change in execution capability.
• Compared to models like “GPT-5.5” and “Opus 4.8”, the speaker argues those behave more like tools that need more explicit instruction, while Fable understands project intent and deeper purpose.
• For workflow/testing, Fable is described as more thorough at verification—the speaker claims it can run dozens of tests automatically when making changes.

2) “Restrictions” discussion: tightened classifier, not full coding ban

• Despite fears that coding/debugging would be disabled, the speaker reports that after hundreds of prompts, downgrades to Opus 4.8 happened only a few times.
• Key factor claimed: using Fable via Cursor (API) is less restrictive than using it in the Claude AI web chat, described as “for the masses.”
• The speaker also claims to have built a custom “skill” that reduces rejection rates.

3) Why it was banned (security vulnerability disclosure)

• The speaker attributes the ban to a security vulnerability discovered while handling a prompt (described as “fix this code”).
• Nuance added: the relevant Amazon employees were allegedly contractually obligated to report serious security risks to the US government first, before notifying Anthropic—so the government’s response contributed to the ban.

4) “Safe without being useless” thesis for coding models

• Core argument: safety doesn’t require crippling coding; instead, models should be made smarter and better aligned so they don’t choose harmful outcomes unnecessarily.
• Restrictions are framed as analogous to societal rules: stronger safeguards when capabilities exceed what’s safe.

5) Practical guidance for using Fable after it leaves Anthropic plans

• Recommendation: use Fable as an orchestrator/planner, and run smaller open-source “actor” agents to carry out actions.
• Mentioned budget-conscious candidates (names may be slightly imprecise): “Kimi 2.7,” “GLM 5.2,” “Minimax,” “DeepSeek/Deep Seek.”
• Budget framing: even with a limited monthly budget, you can still use Fable for planning if you orchestrate and push execution to cheaper models.

6) Building “agent systems” that actually do work: Axiom Work tutorial-style pitch

• The speaker describes a desktop app (“Axiom Work,” from Alibaba.com) used locally, claimed to act as a doer (not only a planner).
• Architecture described:
  • Agents: multiple roles (e.g., sourcing/negotiation) plus a model gateway (cheap fast model for groundwork; Claude for reasoning).
  • Plugins: e.g., a sourcing plugin that pulls supplier shortlists from Alibaba.com transaction data, including ratings, MOQ, prices, and lead times.
  • Connectors: e.g., Gmail for negotiation; Instagram for posting marketing content.
  • Channels: e.g., Telegram for task dispatch; scheduled reruns.
• Claims:
  • Sensitive actions (like payments) require explicit approval.
  • The tool supports keeping data off servers.

7) Handling bans with data collection + skills (“prompt engineering” via skills)

• Operational advice while Fable was unavailable/uncertain:
  • Build datasets by checking API access on a schedule (example given: 400 prompts, polled every 60 seconds across multiple VPS instances).
  • When access returns, the dataset fills and can be reviewed later.
• For long-term productivity: build skills (SOP-like reusable instructions) using Fable so weaker models follow better procedures.
• Specific approach:
  • Create a “Fable safe prompt” skill that rewrites prompts to avoid triggering safety-related downgrades.
  • Example issue: the word “secure” may be interpreted as cybersecurity and cause rejection.
  • The skill is claimed to preserve similar outputs while avoiding downgrades.

8) Open-source catch-up strategy (distillation + datasets)

• The speaker expects open-source models to catch up to Fable in:
  • ~2–3 months (optimistic), or
  • ~5–6 months (more realistic).
• Drivers:
  • Distillation: query a stronger model across many questions, then train smaller models on those outputs.
  • Existing public distillation efforts: they mention 100+ Hugging Face datasets already.
• Larger-scale dataset collection/distribution is discussed (but not endorsed), with attention to Anthropic terms-of-service risk.

9) Cybersecurity perspective: restrictions vs power in open-source

• The speaker argues restrictions mainly limit non-technical users.
• They claim capable actors can still use open-source models to create harmful cybersecurity capabilities.
• They also argue the “bug” isn’t unique to Fable (i.e., other models also found the vulnerability), implying it’s not solely a one-model behavior issue.

10) Startup/strategy: design for agents, not humans

• Strong thesis: most software usage will be by agents, not humans clicking through SaaS UI.
• Product implication:
  • Build CLI tools, clean APIs, and documentation so agents can execute tasks.
  • If a product requires website signup and UI navigation, it’s framed as “not for the future.”
• “CEO vs janitor” thinking: the powerful model/agent should orchestrate, not do every trivial task.
• Scaling assumptions:
  • Increase token growth and the rising ability of agents to execute multi-step workflows.
  • Example: improved productivity for tasks like finding talent/location queries via agent loops (claimed 100x productivity multipliers).

11) Self-hosting / sovereignty / business continuity plan

• To reduce exposure to bans/changes:
  • Self-host open-weight models or use multi-provider routing (e.g., OpenRouter-like distribution).
  • Keep data locally to avoid leakage.
• The speaker argues that buying hardware/GPUs and owning weights is an investment that prevents losing competitive advantage if cloud access changes.

12) Model roadmap speculation and pricing stance

• They believe GPT 5.6 will be unblocked soon.
• Pricing stance:
  • The speaker says they would pay double for Fable due to its being a step-change.
  • Competition and upcoming releases (including GPT 5.6) may pressure pricing.
• Benchmarks/evals are acknowledged but framed as less informative than real-world usability.

---

Main Speakers / Sources

• Primary speaker: the YouTube guest/host persona discussing Fable 5, Cursor, Claude Code, agent orchestration, dataset distillation, and self-hosting (no specific name provided in the subtitles).
• Mentioned third parties:
  • Anthropic / Claude team (including Dario Amodei)
  • Amazon employees (unnamed)
  • OpenAI (GPT 5.6 mentioned; “OpenAI team on vacation”)
  • Elon (referenced regarding deleting/paring down skills)
  • Hugging Face (datasets/distillation)
  • Alibaba.com and Axiom Work (product described)