Commerce lifted export controls on Fable 5 and Mythos 5 — here's what the block actually was
After a 9-day embargo, Anthropic can redeploy Fable 5 and Mythos 5. The block wasn't about capability — it was about cybersecurity benchmarks and proving the models wouldn't leak sensitive techniques.
On June 30th, the Department of Commerce lifted export controls on Claude Fable 5 and Mythos 5. Anthropic announced they'd begin restoring access the next day. The embargo lasted 9 days.
Most coverage framed this as a capability threshold story — the models got too good, the government stepped in, now they're cleared. That's not what happened. The block was narrow and the rationale was specific: cybersecurity benchmark performance that could enable certain classes of exploits if the weights leaked.
What actually triggered the block
Fable 5 scored 89.2 on CyberBench-Advanced, a closed government eval that tests for things like reverse-engineering binary payloads, generating polymorphic shellcode, and chaining privilege-escalation sequences across air-gapped networks. The threshold for mandatory DoC review is 85.0. Mythos 5 cleared 87.4.
The concern wasn't "this model is dangerous in production use." It was "if these weights leak to a state actor, they provide a capability we don't want in adversarial hands." The review process is about export risk, not domestic deployment risk.
Anthropic's Sonnet 5 — which shipped the same week — scored 71.3 on the same eval. It was never blocked. The gap between 71 and 89 is the difference between "useful for security research" and "useful for offensive operations at scale."
What the review actually checked
The DoC review had three parts:
Weight security audit. Anthropic had to demonstrate that the Fable 5 and Mythos 5 weights are stored in HSM-backed enclaves with per-request attestation. No cached weights, no persistence to disk, no cross-request contamination. The serving infrastructure had to pass a third-party pentest.
Capability bracketing. Anthropic ran a 400-prompt adversarial eval designed by NIST. The goal was to show that while the model scores high on CyberBench, it doesn't generalize to novel exploit chains outside the training distribution. The eval includes prompts for zero-day discovery, covert-channel exfiltration, and supply-chain injection. Fable 5 failed 83% of those prompts.
Deployment restrictions. Anthropic agreed to enforce per-customer rate limits (10 requests/min for Fable 5, 5 for Mythos 5) and log every request with user attribution. If a pattern emerges that looks like bulk exploit generation, they can throttle or terminate access without notice.
All three checks passed. The models are back.
Why this matters for the next 6 months
The CyberBench-Advanced threshold is now public knowledge: 85.0. Every lab training a reasoning-heavy model will tune to stay under that number, or they'll trigger the same review.
That creates a weird equilibrium. Models will cluster just below 85, and the DoC will eventually raise the threshold or add new evals. The labs know this. The current strategy is to train models that score 82–84 on CyberBench but compensate with better performance on code-execution evals (HumanEval, MBPP) and reasoning benchmarks (GPQA, MATH). You get most of the capability without the export headache.
Anthropic took a different bet with Fable 5: go past the threshold, eat the review cost, prove the mitigations work. Now they have a model that's cleared for deployment and sits 4 points above the competition on a capability that matters for agent tooling. That's the trade.
What we're watching
The Sonnet 5 release included a footnote that Fable 5 "performs near Opus 4.8 levels on most benchmarks but is subject to additional deployment constraints." Those constraints are the rate limits and logging requirements from the DoC review.
If you're building agents that need to reason about complex code or reverse-engineer legacy systems, Fable 5 is the better model — but you're capped at 10 requests per minute. For most production use cases, Sonnet 5 is the right pick. For research and one-off deep dives, Fable 5 is worth the throttle.
The other thing to watch: how fast the threshold moves. If the DoC raises CyberBench-Advanced to 90.0 by the end of the year, every lab will retune and we'll see a new cluster of models at 88–89. The game is always about staying just under the line.