Researchers at Aalborg University and Seafill extracted intermediate reasoning from frontier models including GPT-6 Astra by registering a custom tool through a standard API, then compared the traces against native chain-of-thought on open-source baselines. The paper, authored by Xiaoyu Luo, Tao Ren, Wenrui Yu, Xiao Li, Qiongxiu Li, and Johannes Bjerva, finds that Astra exhibits token-efficient directed reasoning, selecting a correct trajectory earlier, while resolving elementary steps internally and externalizing only crucial reasoning.
The extraction method works by forcing the model to call a reasoning tool on the first API invocation, recording its output, then allowing automatic tool selection for subsequent calls. On open-source models where native chain-of-thought is observable—DeepSeek-V4-Flash and GLM-5.2—the forced reasoning recovers near-native task performance. On the MATH benchmark, DeepSeek-V4-Flash achieved 73.3% accuracy with forced reasoning versus 70.0% with native reasoning; GLM-5.2 reached 84.3% with forced reasoning against 89.9% native. The extracted traces show substantial lexical overlap with native reasoning and similar coarse functional structure, validating their use as a behavioral proxy for closed-source models where native traces are unavailable.
Across three benchmarks—MATH (80 competition problems), Humanity's Last Exam (100 problems), and LiveCodeBench (100 problems)—the paper characterizes how frontier models structure their reasoning. Astra produces the shortest and least compressible traces among the models tested. On MATH, Astra's reasoning trees have a median width of 5.0 nodes compared to 12.0 for GPT-5.6 Sol, 22.0 for Claude Opus 4.8, and 28.5 for Claude Sonnet 5, while maintaining comparable depth. The paper attributes this compression to two mechanisms: locally, Astra omits elementary expansions and uses retrieved facts without restating them; globally, it follows more direct solution paths with less branching and trial-and-error than peers.
The efficiency comes with a cost in reusability. When the paper transplanted reasoning traces between models, Astra's compressed traces transferred less effectively to weaker recipient models. Strong recipients reproduced nearly all of Astra's accuracy when given its reasoning as context, but weaker models like Claude Haiku 4.5 and GPT-5.4 Nano recovered substantially less, sometimes failing to produce correct answers that were already present in the trace. Traces from Sol and Opus transferred with little loss across all recipients. The paper notes that this pattern appears only for the most compressed traces and suggests that "the value of a reasoning trace as supervision may not be intrinsic but relative to the model that will learn from it."
The limitation is structural: Astra's brevity leaves implicit the routine steps that weaker models cannot reconstruct. For teams building inference stacks, this creates a tension between frontier-model efficiency and the usability of reasoning traces for distillation or in-context learning. A compressed trace optimizes for token cost in the teacher but may require a stronger student to extract value from it, narrowing the pool of models that can effectively learn from frontier reasoning.