MatIntellect

⚡️ A few weeks ago Anthropic warned that the new model was too powerful for public access. On June 9 they released it - in a modified form. Claude Fable 5: the first publicly available Mythos-class model

What is Claude Fable 5:
This isn't Opus Plus - it's a different class. Fable 5 = Mythos 5 under the hood, but with an extra layer of classifier filters on top. When the model detects a request about cybersecurity, biology, chemistry or distillation - the request automatically gets routed to Opus 4.8 and you see a notification. It triggers in less than 5% of sessions. In alignment testing: Fable 5's level of undesirable behavior is on par with Opus 4.8 - same base weights, same numbers. The name Fable - "story" in English - Anthropic explains by the fact that the model builds complex narratives and solves tasks that need long multi-step reasoning. Not one prompt-response, but an extended process with hours of work ahead

Benchmarks - what the tests showed:
🟢 FrontierBench (Cognition coding eval) - best result among all publicly available models
🟢 Core analytics - the first to break 90%, plus 10 points over Opus 4.8
🟢 Frontier physics - 36 hours vs 4 days for GPT-5.5, and with a third of the reasoning tokens

Key detail: the longer and harder the task, the bigger Fable 5's lead over the competition. It's a different kind of reasoning under the hood - the model holds the thread of thought differently. On short answers the gap is less noticeable, on multi-iteration tasks it's a whole other level. Anthropic points this out specifically in the announcement: the harder the task, the bigger the advantage

How to try it:
💎 Pro/Max/Team/Enterprise - free access until June 22
💎 API - $10/M input tokens, $50/M output. Half the price of Mythos Preview
💎 Prompt caching - the 90% discount stays, same as for all Claude models

🛡 Full Mythos 5 - only through Project Glasswing: a government program for cyber defenders and critical infrastructure operators

Mythos 5 - the same base model but without the cybersecurity filters. According to Anthropic: the most powerful model in the world today in terms of cyber capabilities. Available exclusively through Project Glasswing in partnership with the US government. Bloomberg noted: for the first time an American AI lab is coordinating the release of a frontier model through a government program from day one - not after the fact through a partnership, but as the primary distribution channel

The same idea at the level of my agents

Splitting access by trust level isn't just for model makers. I do the same at the agent level. My personal agent Jes has been running for 134 days and handles 10 scheduled jobs on its own: checks, reports, reminders. But deleting data, spending money, messaging clients and publishing sit behind a different gate and always wait for my yes. One model, two levels of trust

At the clinic the levels are stricter still. The owner's voice agent opens only with a password and doesn't answer strangers. The client bot answers only from the clinic's documents, about 114 price items and 48 services, and hands anything not in them to the owner as a card. The clinic runs 6 agents in total, each with its own range of what it may do without a person. That's the architectural answer to safety: don't hope the model will hold itself back; put gates in advance where a mistake is costly

🎯 It's not about the filters - Anthropic split one set of model weights into two access points with fundamentally different levels of trust. Fable 5 for everyone, Mythos 5 only through the government. That's an architectural answer to the question "how do you release a dangerous model": don't restrict one, build two versions with different entry gates. For now all of this rests on trust in Anthropic and in the government partnership. The question is how long this setup holds - and what happens when someone reproduces the same capabilities without those gates

Share:

No comments yet

Leave a Comment

Fields marked with an asterisk (*) are required