Fable 5 is back — but the two-week gap already made its point

On June 12, 2026, the US government forced Anthropic to pull Claude Fable 5 and its safeguards-lifted sibling Mythos 5 worldwide, overnight, under a national-security order. On June 27, the same government lifted the ban. Anthropic is restoring access now, vetted US organizations first, with a broader rollout in the coming weeks, and Fable 5 is back on the benchmark, reclaiming the 95.0% SWE-bench Verified score that made it the class leader before the order pulled it.

What restoration actually means today

• Vetted-first rollout, not a flip of a switch. If you weren't in the first wave of US orgs Anthropic is re-enabling, you're not back yet. Check your own access before you plan around it.
• Mythos 5 is a separate decision. The safeguards-lifted sibling stays a restricted rollout with no published score of its own: its SWE-bench cell still reads n/a. Fable's return doesn't imply Mythos follows the same path or timeline.
• Pricing and context are unchanged. $10 / $50 per million tokens, 1M context: the model coming back is the same model that left, not a discounted or degraded relaunch.

Fifteen days doesn't erase the lesson: it proves it

The reaction to the withdrawal argued that a hosted, closed, frontier model can vanish worldwide by government order with no notice, and that the only real insurance is provider abstraction plus an open-weight floor validated before you need it. The ban lifting in two weeks doesn't undercut that argument. It sharpens it. Fifteen days is exactly long enough to hurt any team that had hard-coupled a critical workflow to one model, and exactly short enough that most people will shrug it off as a blip instead of fixing the actual exposure. A risk that resolves quickly is still a risk. The fire drill ended well this time; that's not a reason to disconnect the smoke detector.

Everything about a stack hard-coupled to one closed, hosted provider having a single point of failure no SLA covers still holds. Fable coming back doesn't change the architecture — it just means the outage was fifteen days instead of indefinite.

If the withdrawal made you audit your provider abstraction and validate a local fallback, keep both: they cost nothing while Fable is available, and they're the reason a repeat of June 12 is a shrug instead of a scramble next time. If it didn't, this is the reminder to actually do it, not the all-clear to skip it.

What changes on the benchmark

Fable 5's row returns at its pre-withdrawal figures. That's a restoration of a number already measured and published before the order pulled it, not a fresh re-verification. Mythos 5 stays listed with swe=n/a, unchanged. Nothing else on the benchmark moves because of this: GLM-5.2's provisional figure and the GPT-5.6 family's n/a rows stand exactly where the withdrawal piece left them.

The model is back. The reason to have a plan that doesn't depend on any one model being back (export controls included) hasn't gone anywhere.