Registering one custom tool makes closed models externalize their hidden reasoning.

arXiv 2609.26637 registers a simple tool through standard function-calling APIs and gets frontier models, GPT-6 Astra among them, to write out intermediate reasoning. The authors validated against native CoT on open models first, and the extracted traces match native reasoning performance on math, science and code. Tool schemas are an unintended channel into reasoning that vendors chose not to expose.