Anthropic’s Fable 5 launches, safer Mythos power for users
AI
Fable
Anthropic
TLDR: SAN FRANCISCO—Anthropic launched Claude Fable 5 for subscribers, a safer Mythos level model with stronger cybersecurity guardrails and better vision, starting now and costing extra after June 22.
Key Takeaways:
- Mythos Preview rumors in April promised internet breaking vulnerability hunting, but Anthropic kept early access tight under Project Glasswing.
- Claude Fable 5 is now in phased subscriber rollout through June 22, then it requires usage credits and developers pay $10 per million input tokens and $50 per million output tokens.
- Anthropic says Fable 5 logged zero harmful single turn cyberattack planning under 30 jailbreak methods, while routing most biology and chemistry questions to Opus 4.8.
This is Anthropic doing the responsible thing in a very unglamorous way: it is shipping capability first, then charging and gating it while measuring risk in the open. The window is short on purpose, because this model eats compute like it is trying to win a staring contest with your budget.
This is Anthropic doing the responsible thing in a very unglamorous way: it is shipping capability first, then charging and gating it while measuring risk in the open. The window is short on purpose, because this model eats compute like it is trying to win a staring contest with your budget.
Q&A
What happens when Fable 5 runs out during the June 22 window?
Users will need usage credits to keep going, and token based billing for developers will likely steer teams toward prompt strategies that reduce output length.
Why does Anthropic route biology and chemistry questions away from Fable 5?
By separating domains, Anthropic can contain high risk topic behavior while still letting engineers and general knowledge workers benefit from the main model.
If Mythos 5 needed government review pressure, what does that imply for future releases?
It suggests Anthropic could face changing policy requirements, which may force more staged access or tighter scope even after a model is technically ready.
How might the vision improvements change what “software help” looks like?
Teams can share screenshots, diagrams, and UI flows for debugging and planning, which can shorten feedback loops but also raises the stakes for prompt and data handling.
Why is the cybersecurity compliance metric framed around single turn requests?
It hints Anthropic is focusing on preventing direct exploit planning behaviors, but multi turn workflows and agentic tools could still change the risk profile later.
No comments yet. Be the first to share your thoughts!