Anthropic said it will make its frontier model "Claude Fable 5"—previously paused after talks with the US government—available again globally starting the day after July 1, paired with new classifiers that block a wider range of cybersecurity tasks.
July 1 · Anthropic
Claude Fable 5 Returns Worldwide — With Tighter Cyber Guardrails
After a jailbreak report triggered a full suspension under US export controls on June 12, Anthropic is redeploying its frontier Mythos-class model globally with new classifiers that block a specific bypass technique — at the cost of more false positives.
99%+
of the reported bypass technique now blocked
<5%
of sessions fall back to Opus 4.8
~2 wks
offline during government talks
Suspension Timeline
JUN 9
Launch — general availability with strict guardrails
→
JUN 12
Access cut for all users — jailbreak report + export controls
→
JUL 2
Redeployed globally with new classifiers
Token Pricing — Input vs Output
per 1M tokens · output costs 5× the input
Context
1M tokens
up to 128K output
Knowledge cutoff
Jan 2026
API: claude-fable-5
Free through
Jul 7
up to 50% of weekly limits
Praise
Top scores on FrontierCode, Hebbia Finance & ViBench
"Senior research scientist grade" — one-shots complex tasks
Beats Opus 4.8 on large codebase migrations & hypothesis generation
Criticism
Overly sensitive filter downgrades even shopping lists & cake recipes
Single prompts can burn a large share of usage limits
New classifiers expected to raise false positives further
What comes next
An industry framework to grade jailbreak severity — with Amazon, Microsoft and Google.
Anthropic pledges stronger pre-deployment testing, information sharing and joint US-government research. The looser-guardrail Mythos 5 returns on a limited basis for US-based organizations only.
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…