GPT-6 Astra's hidden reasoning extracted via API tool trick

Registering a custom tool through a standard API feature is enough to make closed-source frontier language models externalize the reasoning traces they normally hide, including GPT-6 Astra's. The authors first validated the technique against native chain-of-thought on open-source models before extending it to closed systems.

The paper, "Capable yet Parsimonious: Extracting and Characterizing Hidden Chain-of-Thought in Frontier Models," reports that "the extracted reasoning matches native reasoning performance and substantially outperforms no-reasoning baselines, across competition mathematics, science, and code generation." The authors flag their own hedge up front: "these traces may reflect post-hoc rationalization rather than genuine reasoning."

On GPT-6 Astra specifically, the paper says the model "exhibits token-efficient directed reasoning, selecting a correct trajectory earlier, while resolving elementary steps internally and externalizing only crucial reasoning." Across the models studied, the authors report "systematic differences in how models externalize, compress, and organize reasoning" along three axes: token efficiency, reasoning-step types, and induced reasoning trees.

The authors position the contribution as characterization rather than a safety intervention, describing the work as "a behavioral lens on frontier-model reasoning beyond benchmark scores." The abstract publishes no per-model efficiency numbers, does not name which other closed-source systems were probed alongside Astra, and does not say whether providers were notified before publication.