Запрет Claude Fable 5 для иностранцев: технический разбор и план миграции

Trigger: EAR directive от Commerce Secretary Howard Lutnick → Anthropic CEO Dario Amodei, 12.06.2026. Constraint: block access for any foreign national (deemed export), including foreign Anthropic employees. Anthropic API layer lacks real-time citizenship verification → compliance strategy = global endpoint disable. Latency compliance→shutdown: ~90 min. Affected endpoints: claude-fable-5, claude-mythos-5. Unaffected: claude-opus-4-8, claude-sonnet-4-6, claude-haiku-4-5.

Release: 2026-06-09. Tier: Mythos-class (above Opus). Use case: long-horizon agentic workloads — multi-day code migration, autonomous research pipelines, multi-stage document analysis.

Mythos 5: identical architecture, safety classifiers removed. Distribution: Project Glasswing partners only (critical infrastructure, cybersecurity). Both model IDs subject to EAR directive.

• 2026-06-09: GA Fable 5 + Mythos 5 (Glasswing). Channels: Claude API, AWS Bedrock, Vertex AI, Microsoft Foundry.
• 2026-06-12 evening US: EAR directive issued. Requirement: export license for any foreign national access.
• 2026-06-12 +90min: Anthropic public statement + global disable.
• 2026-06-15: Z.ai ships GLM-5.2, explicitly positioned as Fable 5 alternative.

Layer 0 — DoD demand: unrestricted Claude access for all lawful military purposes. Anthropic deny list: (1) mass domestic surveillance, (2) fully autonomous weapons. Layer 1 — March 2026: Defense Secretary Pete Hegseth designates Anthropic supply chain risk (first US company). Litigation: CA preliminary injunction vs DC Circuit denial of stay. Layer 2 — June 2026: Commerce EAR directive, days after confidential SEC IPO filing. Official BIS rationale: claimed Fable 5 jailbreak → cyber/biosecurity NS concern. Anthropic counter: capability exists in GPT-5.5, DeepSeek V3.

Penwell Law / CSIS: directive text requires export licenses for foreign nationals — does NOT explicitly mandate global API disable. Anthropic decision tree: no real-time citizenship filter at API layer → chose global blackout as compliance path. Open question: citizenship verification + block unverified users as narrower compliant alternative. Precedent established regardless: US admin can force commercial AI model offline within hours.

Tier 1 — Anthropic-internal: claude-opus-4-8. Min migration cost, same API surface.

Self-host regions (non-US jurisdiction): Hetzner DE, OVHcloud/Scaleway FR, AWS eu-central-1/eu-west-1. Local inference on Mac: DeepSeek Metal guide.

Phase 1 — audit: grep codebase for claude-fable-5, claude-mythos-5.

import os

PRIMARY_MODEL = os.environ.get("AI_MODEL_PRIMARY", "claude-opus-4-8")
FALLBACK_MODELS = os.environ.get(
    "AI_MODEL_FALLBACKS",
    "gpt-5.5,mistral/mistral-large-latest"
).split(",")

from litellm import completion
from config import PRIMARY_MODEL, FALLBACK_MODELS

def infer(prompt: str) -> str:
    resp = completion(
        model=PRIMARY_MODEL,
        messages=[{"role": "user", "content": prompt}],
        fallbacks=FALLBACK_MODELS,
        num_retries=2,
        timeout=120,
    )
    return resp.choices[0].message.content

import time
from litellm import completion

MODELS = ["claude-opus-4-8", "gpt-5.5", "mistral/mistral-large-latest"]

for model in MODELS:
    t0 = time.time()
    try:
        completion(model=model, messages=[{"role":"user","content":"ping"}], max_tokens=5)
        print(f"OK {model} {time.time()-t0:.2f}s")
    except Exception as e:
        print(f"FAIL {model} {e}")

• Multi-provider architecture: Anthropic primary + Mistral EU hot standby + self-hosted Tier 3 for critical workloads.
• Deemed export audit: map foreign national employees → model API access paths.
• MCP/Skills versioning: MCP Server guide, Hermes Skills, Cursor Skills.

Subscription: prefer monthly billing; 3-month eval before annual; calendar renewal dates; read refund policy (Anthropic refunded 2026-06-09..14 as exception).

Prompt asset management: export prompts to local storage; document capability requirements not model names; backup .cursor/rules/, SKILL.md, MCP configs to Git.

Alert pipeline: primary sources (Anthropic blog, BIS.gov), Google Alerts, Hacker News. Incident landed Friday evening — act within hours not days.

No SPOF: primary tool + tested backup + free tier familiarity (Claude/ChatGPT/Gemini).

• Precedent: cloud API access = controlled export (dual-use tech parity).
• Anthropic IPO: market confidence hit post-SEC filing + shutdown.
• Vendor lock-in risk: political dimension now first-class in enterprise AI strategy.
• Open source acceleration: GLM-5.2, Qwen3, DeepSeek benefit from trust deficit.
• Paradox: export control accelerates open alternatives it targets.

1–6 months: Anthropic citizenship verification R&D; Pentagon litigation ongoing; AI Diffusion Rule legal status contested (GAO May 2026).

6–24 months: systematic US AI export framework (chip-regime analog); EU AI sovereignty → Mistral enterprise adoption; open-weight frontier parity; citizenship verification as standard onboarding.

1. 01
   Audit + migrate: replace claude-fable-5 → claude-opus-4-8. Ref: assistant comparison.
2. 02
   Deploy LiteLLM router: provision cloud Mac via NUKCLOUD console. Pricing: tseny.html.
3. 03
   Version MCP + Skills: Git-track configs — MCP from scratch, Hermes Skills.
4. 04
   Enable EU fallback: Mistral Large 2 hot standby. Validate routing via OpenRouter trends.
5. 05
   Evaluate local open-weight: DeepSeek/Qwen on 32GB+ Mac — Metal inference. Cost optimize: June deals.
6. 06
   Production harden: health checks, deemed export policy, BIS alerts. Reserve capacity: zakaz.html.

Agent workloads on laptops fail on sleep, bandwidth, sudden API shutdowns. NUKCLOUD bare-metal Mac provides tenant isolation + 24/7 uptime for multi-provider routing and local inference.

• Anthropic official statement
• CSIS analysis
• Penwell Law legal analysis
• LiteLLM
• GAO AI Diffusion Rule
• NBC News