Fables 5 has returned three weeks after the US government ordered its blackout. It returns with several asterisks: it only allows 50% of weekly usage for paid plans, and it brings a more aggressive security filter along with a commitment that binds Anthropic to share with the government every model that comes after it, including future ones.
The official origin is a reproducible failure in models much inferior to Fable. But what remains after solving it weighs more: early government access before each launch and dedicated computing to audit the models.
In detail. On June 12, the US government acted after learning of a report from Amazon researchers that Fable 5 had identified software vulnerabilities by bypassing its safeguards. Anthropic replicated the finding with less powerful models (Opus 4.8, GPT-5.5 and Kimi K2.7) and everyone got it the same. On June 30, the restriction was lifted.
Of course, the new classifier notifies the user when it blocks a request and redirects it to Opus 4.8.
Between the lines. Anthropic admits that this classifier blocks more harmless requests, especially in programming and debugging code, its main use. The safety margin itself, the company says, is “far greater than in any previous release.”
Alberto Romero, in The Algorithmic Bridgereads it like this: if the model fails due to excess caution in its most common use, the real ceiling of what a user can squeeze out remains at the level of Opus 4.8 either GPT-5.5although the model below is more powerful.
Why is it important. Anthropic grants early access to models that “materially advance the capability frontier” in national security, and assembles internal teams for government priorities.
The government reserves the right to reimpose the license “if circumstances change”, but without defining which ones. Since the premise of this industry is that a larger model develops new capabilities overnight, almost any advancement fits that phrase.
The context. Anthropic has been collaborating on an ad hoc basis with US agencies such as the Treasury and the National Cybersecurity Office for almost two years. What is new is that this work becomes a permanent protocol, and that Fable and Mythos They are the first to pass through it.
Anthropic has been chasing OpenAI for years, but in 2026 it has taken the lead on several fronts, in the heat of Claude Code, Claude Cowork and the discovery of the quality of their models’ responses by those who had not left ChatGPT. Accepting this tutelage now, when they finally have something to lose, is a gamble: they prefer a guarded border to a race that neither they nor their government fully control.
And now what. What needs to be monitored is whether the classifier improves its accuracy in code without losing a strong hand, and whether OpenAI or Google end up signing something similar. If that happens, Fable’s return will no longer seem like an isolated incident.
Featured image | Anthropic
GIPHY App Key not set. Please check settings