A new technical work has demonstrated that it is possible to fully read the reasoning tokens produced by Claude and GPT models during inference. The study doesn’t stop at the mechanism: it offers concrete examples, and those samples expose three dynamics that change how we must look at proprietary models.

The most revealing is the suspicion of benchmaxing. When faced with a problem from the AIME benchmark, Claude’s reasoning showed that the model knew both the question and the answer by heart, before it even began to “think”. If the scores we see on leaderboards partly reflect data already encountered during training, the real gap with open source models could be much narrower. There is no secret sauce: at the bottom, it’s data, compute, and engineering, and some benchmark numbers may be overstated.

The second signal is about overthinking. Anyone working with self-hosted models frequently encounters chains of thought that loop on themselves, with words that appear meaningless or unnecessary convolutions. It’s often seen as a flaw of less sophisticated stacks, but the reasoning extracted from frontier models shows the phenomenon is just as pervasive there. It’s a finding that advises recalibrating expectations when debugging on-premise pipelines: imperfection is not a sign of technical inferiority, but a constant of LLM inference.

The third effect is geopolitical and concerns distillation. For some time, many have discussed whether actors like China exploited this very opening to distill frontier models through APIs, extracting reasoning tokens to train competing models. Now that the gap has been closed, that channel of forced intellectual property transfer slows down. It’s a reset that benefits closed labs, but at the same time makes competition fairer for those developing in open source: the ability to produce competitive models will have to rely on their own innovation, not on the covert copying of chains of thought.

For those evaluating in-house deployment, the episode adds concrete evidence to the data sovereignty dossier. When an API model also returns its reasoning, even if theoretically invisible, it creates an unexpected exfiltration surface. Companies processing sensitive information — in legal, healthcare, or financial contexts — cannot ignore that trade secrets could travel, albeit in the form of computational thought, through uncontrolled infrastructure. An on-premise architecture, or at least an air-gapped network perimeter, becomes a non-negotiable prerequisite to neutralize this risk.

It’s not so much the technical discovery that weighs — the gap was probably known to labs and state actors — but the effect of public disclosure. The excess of trust in benchmark metrics and in the opacity of proprietary models clashes with evidence: AIME scores and similar are not completely unbiased, and a model’s true reasoning remains a box of which today we saw one more piece. Closing the gap doesn’t erase that piece, but makes it harder to spy on the box.

Open source, despite appearances, may not be so far behind. And the next generation of self-hosted models will have one more reason to claim transparency: you can inspect every reasoning token because the model runs under your control. That’s an argument that in compliance departments will weigh more than any paper.