Advisor Room · Fort Lauderdale, FL · July 2026
Fable 5 &
Mythos 5
The Masterclass — Relaunch Edition
~35 minutes24 slidesRedeployed July 1
History · The models · The cyber plan · The shutdown · Redeployment · How to recover
Style can be copied.
Skill can't.
Constitutional AI gives the model a 'constitution'—principles it can use to reason about its own responses: helpful, harmless, honest.
- 2021: Dario Amodei (VP Research) and Daniela Amodei (VP Operations) leave OpenAI along with about nine colleagues.
- The reason: concern that Microsoft's $1B investment was shifting OpenAI's focus away from safety.
- They wanted a lab where safety research and product were the same thing—not separate teams.
The frontier paradox: they believe they may be building dangerous technology—and build it anyway, because a safety-focused lab at the frontier is better than leaving that ground to others.
- Claude 1
- Mar 2023
Constitutional AI in production.
- Claude 2
- Jul 2023
100K context window.
- Claude 3
- Mar 2024
Haiku, Sonnet, and Opus.
- Claude 3.5
- Jun 2024
Best coding model in the world.
- Claude 4.x
- 2025
Agentic reasoning and Claude Code.
- Fable 5
- Jun–Jul 2026
Suspended three days after launch; redeployed globally July 1.
- Fable 5 and Mythos 5 launched June 9 and were suspended June 12. The export controls were lifted June 30, and access began returning July 1.
- The gap between Fable and Opus splits in two: raw capability in the weights, and working-style texture. Only the texture transfers with prompts.
- The session explains the models, the suspension and redeployment, and how to close the recoverable half of the gap when a workload does not require Fable.
- claude-fable-5
- Fable 5
Restored globally July 1 on Claude Platform, Claude.ai, Claude Code, and Claude Cowork. Classifiers active.
- claude-mythos-5
- Mythos 5
Restored for a set of US organizations. Broader Glasswing access remains under coordination.
- claude-opus-4-8
- Opus 4.8
Available now and unaffected by the suspension.
The clean mental model: Mythos is the unrestricted engine. Fable is the same engine wrapped in classifiers for public release.
- Context
- 1M tokens
Opus 4.8: 200K
- Max output
- 128K
Opus 4.8: 32K
- Input
- $10 / 1M
Opus 4.8: $5 / 1M
- Output
- $50 / 1M
Opus 4.8: $25 / 1M
- Retention
- 30 days
No zero-data-retention option
- Availability
- Restored
Eligible Claude plans moved to usage credits after July 7; Mythos remains limited to a set of US organizations
Compliance note: if zero data retention is a contractual or regulatory requirement, neither model is an option for that workload—period.
- During Mythos Preview, it identified an extraordinary number of zero-day vulnerabilities across major operating systems and browsers.
- It did not just find them—it actively exploited them. Wall Street and government intelligence took immediate notice.
- Anthropic itself lobbied for Mythos to be regulated as a cyberweapon.
The dual-use problem: in the wrong hands, exploit capability could target banking systems, power grids, and hospital networks.
Before releasing Mythos publicly, Anthropic gave it to vetted cybersecurity companies so they could find and patch the same vulnerabilities Mythos could exploit.
- Mythos can find and exploit vulnerabilities better than any human team.
- Security companies can use that same power to patch those vulnerabilities.
- By public release, the known attack surface should already be hardened.
- Project Glasswing uses a separate vetting process arranged through Anthropic, AWS, or Google Cloud account teams.
For security tooling, Mythos through Glasswing remains the access path. Anthropic restored it for a set of US organizations and is coordinating broader domestic and international access.
- Door 1
- Mythos
Vetted access, full exploit capability, no safety classifiers.
- Door 2
- Fable
Public access, same model, classifiers block offensive cyber output.
That was the central safety concern. Anthropic's later review found that the reported technique did not expose unique Mythos-level capabilities.
- cyber: malware, exploit development, or attacking specific targets. Benign pen-testing, CTF, and education can also trip it.
- bio: dangerous lab methods or pathogen enhancement. Beneficial life-sciences work can also trigger it.
- frontier_llm: assistance developing competing AI models under Anthropic commercial terms.
- reasoning_extraction: requests to reproduce internal reasoning. Use adaptive thinking instead.
The redeployed Fable uses an improved classifier that blocks the reported technique in over 99% of cases. Blocked requests are disclosed to the user and sent to Opus 4.8 instead.
- June 9
- Launch
Fable public; Mythos limited to Glasswing.
- June 12
- Suspended
Immediate export controls make real-time nationality verification impractical.
- June 30
- Controls lifted
Anthropic announces a constructive resolution and updated safeguards.
- July 1
- Access restored
Fable returns globally; Mythos returns for a set of US organizations.
- Amazon researchers found a method that prompted Fable to identify software vulnerabilities and, in one case, demonstrate an exploit.
- Anthropic confirmed that less-capable models could identify the same vulnerabilities and that every model it tested could produce the same exploit demonstration.
- The reported technique involved routine defensive cybersecurity work and did not expose unique Mythos-level cyber capabilities.
- No universal jailbreak for Fable 5 had been discovered when Anthropic published the update.
Anthropic classifies the reported behavior as a borderline case inside its intentionally conservative safety margin.
- Anthropic trained an improved classifier specifically targeting the behavior in the Amazon report.
- The new classifier blocks the technique in over 99% of cases.
- Researchers from the US Department of Commerce's CAISI tested the prior and new safeguards and agreed that they are extraordinarily strong.
- The tradeoff is more false positives during legitimate coding and debugging work.
When Fable blocks a request, the user is notified and the request is sent to Opus 4.8.
- Anthropic now confirms that Amazon researchers found the reported bypass and that Amazon participated in the subsequent review.
- The post does not confirm earlier reporting about direct contact between Amazon's CEO and the administration.
- The post does not address earlier claims about a China-linked group accessing Mythos.
- Anthropic, Amazon, Microsoft, Google, and other Glasswing partners are drafting a shared jailbreak-severity framework.
The proposed framework scores capability gain, breadth, ease of weaponization, and discoverability.
- Runs longer. Hard tasks run for minutes; autonomous runs extend for hours.
- Plans more. One fix: 'when you have enough to act, act.'
- Follows instructions more tightly. One brief instruction beats enumerating behaviors.
- Dispatches subagents more readily. Guide when to delegate versus work inline.
- Old Opus-era prompts can be too prescriptive and actively degrade Fable output. Migration is partly subtraction.
Claude Code stores every session as JSONL on disk: a real behavioral record, not estimates.
- Median words
- 47 → 18
Opus 4.8 → Fable 5
- Mean words
- ~100 → ~99
Essentially equal
- Tool-to-prose
- 1.41 → 3.91
Fable does 2.7× more work per word
- Opening
- 'I'll' → 'Done'
Narration gives way to results
It's distribution, not conciseness. The means are equal; the 1.41 versus 3.91 ratio quantifies the difference.
- Half 1
- Texture
Verbosity, hedging, openings, and prose-to-work ratio. Steerable with prompts and hooks.
- Half 2
- Capability
First-shot correctness, long-horizon autonomy, vision, and multi-step reasoning. It lives in the weights.
No prompt recovers raw capability. That needs task structure and multi-LLM orchestration.
fable-mode enforces staged execution on large tasks: decompose before acting, delegate in parallel, verify with failable checks, and critique before delivery.
- Stage mapping: write the full plan with numbered stages and expected outputs before acting.
- Parallel delegation: give each delegate a task, deliverable, storage location, and prior context.
- Failable verification: each stage needs a check with real fail conditions.
- Self-critique: review as a skeptical outsider, then fix or disclose real weaknesses.
- Default
- fable-mode
Runs inline on the current model.
- Sonnet
- fable-sonnet
Strong reasoning at lower cost than Opus.
- Haiku
- fable-haiku
High-volume or cost-sensitive work.
The ceiling: this imposes procedure, not capability. Treat it as a checklist, not a transplant.
- Adaptive thinking: internal reasoning with only the conclusion output.
- Classifier: the gate layer between Mythos and Fable; it returns stop_reason: 'refusal' on a hit.
- Effort: the API parameter controlling how much compute the model spends on a task.
- Governor: an eight-rule behavioral contract in CLAUDE.md that steers Opus toward Fable's working style.
- Leak-test harness: reads session logs and measures median words, tool-to-prose ratio, and opener style.
- Project Glasswing: the controlled rollout program that gives security companies Mythos access first.
- Re-injection hook: reprints the governor block every turn to counter salience decay.
- Tool-to-prose ratio: the amount of tool use relative to explanation text.
- Two-doors design: Fable's classifier plus Mythos's restricted access; both must hold.
The two-sentence model: Mythos is the unrestricted base model. Fable is Mythos plus a classifier layer that blocks four refusal categories.
- A CLAUDE.md block recovers working-style texture—which, for bounded daily work, was most of what people were missing when Fable was pulled.
- The measurement discipline outlives the outage. Now that Fable is available again, the harness can tell you whether a workload actually needs it.
- The responsible-disclosure strategy—security companies first, patch the holes, then release—is the right model for powerful dual-use AI.
- Whether the suspension was proportionate is a separate question.
Advisor Room · Fort Lauderdale, FL · July 2026 · cal.com/fractional-cto/discovery-call