The deal Anthropic is negotiating with Washington reopens an uncomfortable question: how do you control the export of something that travels in a file

Export restrictions were conceived with physical goods in mind: machines, components, chips. That is why the episode reported by Bloomberg is so revealing, according to which Anthropic and the Trump administration would be in the final stretch of an agreement to lift the controls weighing on its two most advanced models, Fable 5 and Mythos 5. It is worth being precise about the attribution: the information comes from knowledgeable sources cited by the outlet, and the article itself does not detail the technical terms agreed upon or the markets affected.

According to those sources, Commerce Secretary Howard Lutnick would be the main driver of the negotiation and would be resolving the security concerns that gave rise to controls he himself imposed. The lifting would be conditioned on the approval of the various administration bodies, which points to an interagency validation process that is still open. We are not, therefore, facing a done deal, but rather an ongoing negotiation whose timelines have not been made public.

What the case clearly illustrates is the growing tension between two legitimate goals of the U.S. government: boosting domestic AI development while, at the same time, preventing frontier technology from reaching the hands of geopolitical rivals. Applying export controls to specific models of a private company is a far-reaching measure, and the fact that the company itself is negotiating to reverse them suggests it perceives those restrictions as a brake on its international competitiveness. It is a reasonable reading, although it should not be taken as proven beyond what the sources indicate.

Here lies the real conceptual knot. Unlike a chip, a model's weights are information: they can be copied and transmitted digitally, which makes any export control harder to implement and, above all, to verify. What exactly does it mean to 'not export' a file? The agreement being forged could set a precedent for how frontier software is regulated, and hence its interest beyond the name of the company.

For those who build agentic systems, the implications are tangible. Anthropic's advanced models are commonly used as the reasoning engine in agent pipelines, and an export unlocking would broaden international customers' access to those capabilities. The balance, in editorial terms, is one of cautious optimism: if the outcome is a framework that reconciles national security and technological availability with clear rules, the sector would come out ahead. The unknown —and it should be acknowledged— lies in the details that, for now, remain unknown.