Anthropic executives traveled to Washington, DC on June 15 to meet in person with White House and Commerce Department officials, yet the two sides remain divided over the risks posed by the company's most advanced model, Claude Fable 5.
June 15, 2026 · Anthropic vs. Washington
A record-setting AI model — pulled offline by the government
Claude Fable 5 tops the benchmarks, but Commerce-Department export controls have suspended access since June 12 over fears that jailbreaks could unlock Mythos-level cyber capabilities. Washington talks ended with no deal.
Suspended
Access blocked since June 12 under an export-control directive
<5%
of sessions trigger cyber / bio / chem guardrails (fall back to Opus 4.8)
No deal
June 15 White House talks ended without agreement on risk
SWE-Bench Pro — Fable 5 leads the field
Share of software-engineering tasks solved (%)
The cost question: ~2× the price for a 5–6% gain
Input price per 1M tokens — each block = $2
Output runs $50 / 1M tokens · 1M-token context, 128k output · 90% input discount with prompt caching
What it does well
Stripe migrated a 50M-line Ruby codebase in a single day
Top score on Cognition's FrontierCode
Preferred 80% of the time for research hypotheses
Epoch Capabilities Index 161 — ranks first
The pushback
Safety fallbacks fire even on harmless queries
Heavier token consumption on big tasks
~2× the price for a marginal gain
Sudden suspension disrupted customers
The standoff over Mythos-level capability
GOVERNMENT'S FEAR
Jailbreak prompts could break Fable 5's guardrails and unlock Mythos-level cyber abilities — including zero-day discovery.
ANTHROPIC'S CASE
The risk is overstated; researchers' open letter says capabilities match other leading models — and removing defenders' best tool is unjustified.
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…