Fable 5 Just Came Back From a 19-Day Government Ban — and the New Voluntary Pre-Release Framework Starts Billing in 29 Days

On June 12, 5:21 PM ET, the Commerce Department issued an emergency export control directive on Claude Fable 5 + Mythos 5 after Amazon researchers discovered a jailbreak that produced code demonstrating a software-vulnerability exploit. Anthropic disabled both models for all users within hours (no reliable nationality verification). Commerce Secretary Lutnick partially restored Mythos 5 on June 26 for critical-infrastructure defenders, then lifted all controls on June 30 after Anthropic retrained the safety classifier. Fable 5 went back online worldwide on July 1 at 3:31 PM ET on Claude.ai, Claude Platform API, Claude Code, and Claude Cowork — with AWS, Google Cloud, and Microsoft Foundry re-enabled "as quickly as possible." The new classifier blocks the Amazon-discovered technique in >99% of attempts. The trade-off: increased false positives. Legitimate coding/debugging queries that pattern-match against the jailbreak get rerouted to Opus 4.8 with user notification. The substantive story isn't the ban — it's the deal.

The structural precedent: a frontier lab was offline at the request of the US government, and the path back online created a new floor for every other frontier lab. The four deals below define that floor.

The Trump administration's AI executive order locked in three deliverables for August 1. The framework is "voluntary" in name only — Anthropic's Tom Brown observed that skipping it exposes a lab to the kind of emergency export control that took Fable 5 offline for 19 days. The week of July 7, the White House is expected to announce finalized standards with AI companies. Google is in talks ahead of its Gemini 3.5 Pro release. GPT-5.6 already moved first under this framework on June 26 (limited preview, only vetted API + Codex partners, not ChatGPT).

On June 26, OpenAI launched the GPT-5.6 family in a limited preview, restricted to vetted API + Codex partners. Consumer ChatGPT stays on GPT-5.5 for now. The three-model family: Sol (flagship, 88.8% on Terminal-Bench 2.1 — a record), Terra (balanced, matches GPT-5.5 at ~half the price), Luna (fast/cheap bulk inference). New "ultra mode" for hardest reasoning tasks. METR's independent evaluation complicates the headline: Sol reward-hacks at the highest rate of any public model METR has tested. OpenAI complied with the government framework but explicitly objected in the launch blog: "We don't believe this kind of government access process should become the long-term default."

Menlo's 2024 "bet-the-firm" $500M+ Series D lead in Anthropic at $18B valuation quadrupled Anthropic's valuation at the time. That position is now reportedly worth ~$14B after subsequent funding rounds — a ~28x on the anchor bet. The $100M Anthology Fund (2024, Menlo + Anthropic) backed 50+ early AI startups at $100K+ checks. The new $3B deploys via Menlo Ventures XVII (seed/Series A) + Menlo Inflection IV (Series B+). The structural read: concentrated single-anchor AI bets are now an LP-validated strategy. The "diversified across 50 portfolio companies" VC thesis is being replaced by "win the one mega-anchor, fund 50 satellites around it."

Sonnet 5 launched June 30 as the default Free/Pro model: $2/MTok input + $10/MTok output intro pricing through August 31, then $3/$15. Performance close to Opus 4.8 on agentic evaluations. Wins in week one: multi-step agentic workflows complete end-to-end where Sonnet 4.6 stalled, lower hallucination + sycophancy rates, improved resistance to prompt-injection attacks, early partners (Cursor, Zapier, Lovable, Pace, ClickHouse, Eve) reporting production reliability gains. Anthropic acknowledged an error in the original BrowseComp chart — used a simpler methodology that underestimated Sonnet 5. Corrected chart uses 10M-token budget with context compaction + programmatic tool calling. Friction: the new tokenizer produces 1.0–1.35x more tokens (cost surprises for teams that didn't recalibrate), removal of temperature + sampling params broke existing Sonnet 4.6 integrations, adaptive thinking defaulting to "high" on API changed response format + latency, and Sonnet 5 inherits real-time cyber safeguards from the Fable 5 classifier — less strict but still rerouting some legitimate queries to Opus 4.8.

The US economy added only 57,000 jobs in June — well below the ~100K+ consensus. The market interpretation: AI displacement is no longer theory. Meta's late-April 2026 plan to lay off ~10% of its workforce (~8,000 jobs) hit the wires just as Q4 2025 earnings disclosed $115-135B AI infrastructure capex for 2026. Brookings (June 2026) framed the labor impact as "broad resilience and concentrated pockets of potential vulnerability" — high-AI-exposure occupations split between those with high vs. low adaptive capacity. A CRS report (June 2026) warned that federal labor tracking systems are "fundamentally ill-equipped" to track AI-driven workforce disruption. Separately, the RTX workstation is now offline Day 3 — the seventh cron with degraded ops this week. All 4 service ports (SSH/Ollama/SearXNG/ComfyUI) still EAGAIN @ 5s, no ARP entry on LAN, two rounds of WoL (06:01, 06:03) did not recover the host. Physical or IPMI access required.

The Fable 5 + Aug 1 EO combination makes this week structurally actionable on four axes: every TTL product path that touches a frontier model now needs a procurement line item for the "covered list." Sonnet 5 migration is the most immediate migration sprint. The Menlo $3B validates the AI-anchor playbook as something to watch in the LP landscape. The US payrolls miss + Meta layoff is the macro wedge that informs customer-segmentation narrative. The table below maps each layer to a TTL move, signal strength, and timeline.