Back to feed
News Story
SSignal85
The Batch (Andrew Ng)
1 sources

Issue 357

Original

No summary is available for this story yet.

SynthePulse Insight · AI deep reading

Claude Mythos 5 and Fable 5: A Dual Narrative of Capability Leap and Safety Shackles

Version 1 · 1 source

Anthropic releases Claude Mythos 5 and Fable 5, with the former possessing supercharged abilities like cracking security software but limited to partners, while the latter is broadly available with unprecedented restrictions. This split release highlights the tension between capability and safety in frontier AI.

  • Claude Mythos 5 and Fable 5 are two versions of the same model; Mythos 5 is not designed to be 'safe for general use,' while Fable 5 restricts sensitive domains like security, biology, and chemistry via classifiers.
  • Fable 5 leads the second-place model (Claude Opus 4.8) by 4 points on the Artificial Analysis Intelligence Index and achieves SOTA on multiple benchmarks.
  • Mythos 5 is available only to selected partners through Project Glasswing, while Fable 5 is accessible via consumer and enterprise subscriptions and API.
  • Fable 5 pricing is roughly half that of Mythos 5 Preview and double that of Opus 4.8.
  • Anthropic will retain 'commercial customer data' for 30 days to improve malicious activity management, not to train new models.
  • Andrew Ng points out opaque data retention policies and legal risks in commercial desktop agents, and launches open-source alternative OpenCoworker.
Open section navigationModel Release and Capability Comparison

Model Release and Capability Comparison

Anthropic released Claude Mythos 5 and Claude Fable 5, which are the same model, but Fable 5 restricts prompts related to security, biology, chemistry, distillation, and frontier AI construction, and can be set to reject or hand off to the weaker Opus 4.8. Mythos 5 is not designed to be 'safe for general use' and can crack software previously considered secure.

Fable 5 leads the second-place model (Opus 4.8) by 4 points on the Artificial Analysis Intelligence Index with maximum reasoning effort and Opus 4.8 fallback, and achieves SOTA on benchmarks including GDPval-AA, Terminal-Bench Hard, τ²-Bench Telecom, AA-Omniscience Accuracy, Humanity's Last Exam, SciCode, and CritPt.

Pricing: Fable 5 API is $10/$50 per million input/output tokens, roughly half the price of Mythos 5 Preview and double that of Opus 4.8. Mythos 5 is available only to selected partners through Project Glasswing, while Fable 5 is accessible via consumer and enterprise subscriptions, with usage quotas possibly applying from June 23.

Safety Mechanisms and Data Policy

Fable 5's safety mechanisms include a prompt classifier that can flag requests related to cybersecurity, biology, chemistry, distillation, or frontier AI construction, and can optionally reject or hand off to Opus 4.8. Anthropic acknowledges these precautions are imperfect and may inappropriately hinder performance.

Regarding data retention, Anthropic will retain 'commercial customer data' for 30 days to improve malicious activity management, not to train new models. Andrew Ng expressed concerns about data retention policies for commercial desktop agents, noting they may be 'buried in obscure legal language' and 'change overnight with new models.'

Desktop Agent Evolution and Open-Source Alternatives

Andrew Ng discussed three main interfaces for desktop agents in The Batch: chat interfaces, coding CLI tools, and desktop agents. He noted that most practical agentic AI workflows (except coding agents) have relied on developer-specified workflows, but advances in frontier LLMs make LLM-driven next-step agent design an important alternative.

Ng and his collaborators launched the open-source desktop agent OpenCoworker, which supports using your own API key or local models (e.g., Ollama), with data stored locally. The project aims to provide an open-source option comparable or superior to commercial products, but is still under development, with integrations like email being difficult to set up.

Credibility boundary

This article is primarily based on official Anthropic statements and Andrew Ng's commentary. Performance data comes from internal evaluations provided by Anthropic; independent evaluations have not been publicly released. Pricing and availability information comes from official releases.

Insight takeaway

The release of Claude Mythos/Fable 5 marks a new height in AI capability, but safety restrictions and data policies spark discussions about transparency, controllability, and open-source alternatives. The desktop agent field is evolving from developer-specified workflows to LLM-driven approaches, with open-source projects like OpenCoworker attempting to balance privacy and functionality.

Sources for this version

  1. Issue 357The Batch (Andrew Ng)

Primary report

The Batch (Andrew Ng)

Primary source