The Pulse
GPT-6 Astra Shows Token-Efficient Reasoning in New arXiv Study
A new arXiv preprint finds that GPT-6 Astra produces compact, directed reasoning traces while maintaining strong performance across mathematics, science and code-generation tasks.
AI.info Team ·
GPT-6 Astra produces a median reasoning tree with 27 nodes, compared with 136 for Claude Sonnet 5, in a new study that probes how frontier language models work through difficult problems. The researchers say Astra reaches similar depths with far less branching, revisiting and trial-and-error than the other models tested.
The results come from a preprint by Xiaoyu Luo, Tao Ren, Wenrui Yu, Xiao Li, Qiongxiu Li and Johannes Bjerva, affiliated with Aalborg University and Seafill. Submitted to arXiv on September 22, 2026, the paper introduces a protocol for making intermediate reasoning visible through a client-defined tool, then compares the resulting traces across mathematics, science and code-generation tasks.
Xiaoyu Luo, the paper's first author, wrote:
We find that Astra exhibits token-efficient directed reasoning, selecting a correct trajectory earlier, while resolving elementary steps internally and externalizing only crucial reasoning.
The authors caution that the extracted text may be a useful behavioral proxy rather than a literal record of a model's internal computation. For closed models, native chain-of-thought is unavailable, so the researchers cannot establish that every visible step reflects what happened inside the model.
A tool turns hidden reasoning into recorded text
The protocol, called FORCED-REASONING, registers a tool with a single free-form string argument for intermediate reasoning. Researchers force the model to select that tool first, record the text placed in the tool call, return a content-free acknowledgment and then allow the model to continue choosing between the tool and a final answer.
The tool performs no external computation. Its purpose is to create a visible workspace inside the conversation. The experiments run through OpenRouter, with native reasoning disabled where the model permits it; Astra does not support disabling native reasoning and is tested at its lowest available setting.
Before studying closed models, the researchers test the method on DeepSeek-V4-Flash and GLM-5.2, where native reasoning is available for comparison. Extracted traces on those models achieve near-native task performance, show lexical overlap with native chain-of-thought and display broadly similar functional structure.
That validation supports using the traces as a comparison instrument, but it does not turn them into direct access to hidden computation. The paper explicitly separates behavioral evidence from proof that a model generated the same reasoning internally.
GPT-6 Astra takes a narrower path
The main closed-model comparison covers Claude Opus 4.8, Claude Sonnet 5, GPT-5.6 Sol and GPT-6 Astra. The evaluation uses 80 competition-level mathematics problems, a 100-question subset of Humanity's Last Exam and a 100-question subset of LiveCodeBench.
On the mathematics set, Astra scores 93.8% with the forced protocol against 97.5% with native high-effort reasoning. On Humanity's Last Exam, it scores 32% forced versus 35% native, while on LiveCodeBench it reaches 89% forced versus 92% native. The protocol therefore produces performance close to native reasoning while exposing text that providers normally do not show.
The structural comparison is more striking than the accuracy gap. On the mathematics problems, Astra's median reasoning tree has a width of 5, a depth of 11 and 27 nodes. GPT-5.6 Sol records a width of 12 and 60 nodes; Claude Opus 4.8 reaches 22 and 96 nodes; Claude Sonnet 5 reaches 28.5 and 136 nodes.
Tree depth measures the furthest occupied reasoning step, while width captures the largest lateral expansion and node count records the total number of reasoning nodes. Astra's similar depth but smaller width and size indicate that it follows comparably deep solution paths with fewer detours.
Shorter does not mean less capable
The researchers find that the models use broadly similar categories of activity, including reading, analysis, planning, implementation, exploration, verification and monitoring. Their main difference is how much of that work they externalize and how they arrange it.
Astra often compresses arithmetic checks, algebraic transformations and familiar facts into short statements. The paper describes traces that omit articles, auxiliary verbs and punctuation, using fragments and telegraphic phrases to encode intermediate conclusions. The style appears optimized for compact information transfer rather than a fully articulated explanation for a human reader.
The model also records the shortest and least compressible traces among the systems studied. The authors use lossless zlib compression as a rough measure of redundancy: a higher compressed-to-original ratio indicates text with fewer repeated or predictable patterns. Astra has the highest ratio across all three benchmark groups while producing fewer output tokens.
The findings do not show that Astra performs fewer reasoning operations. Instead, the study says it appears to resolve elementary steps internally and externalize higher-level decisions and checks. The visible trace is sparse, but the researchers identify a similar overall repertoire of reasoning activities.
Compressed traces create a problem for weaker readers
The team also tests whether one model can use another model's extracted reasoning. Each recipient receives a donor trace without the final answer and must solve the problem without tools or native reasoning.
Traces from Sol and Opus transfer with little loss across the tested recipients. Astra's traces behave differently: stronger recipients recover nearly all of Astra's accuracy, while weaker systems recover less. The largest shortfalls appear with Claude Haiku 4.5 and GPT-5.4 Nano.
In some cases, a trace states the correct answer and the recipient still returns a different one. The result suggests that a compact reasoning trace can sit beyond the capability of the model reading it. A student model may need the intermediate steps that a stronger teacher has omitted.
The authors do not test post-training distillation directly, so the study does not establish that compression causes a particular training outcome. Their experiment measures in-context reuse, and donor traces differ in correctness, content and style. They propose a testable follow-up: expanding a compressed trace should help weaker recipients more than stronger ones.
What the protocol does not prove
The study's security implications are limited by its access requirements. The method needs an API that supports custom tools and forced selection of a named tool; systems without those controls fall outside its scope. The researchers say all extraction experiments used benign public benchmarks and did not target personal data, proprietary prompts, credentials or unsafe content.
The paper says the researchers reported their observations by email to security teams at OpenAI and Anthropic and provided code to reproduce the procedure. It also notes that the protocol can expose reasoning-like content even when a provider reports zero reasoning tokens or disables a designated reasoning mode.
That distinction matters. The work shows that controls around a model's official reasoning channel do not necessarily prevent intermediate text from appearing through another API field. It does not show that the text is the model's private chain-of-thought, nor that every exposed trace faithfully represents the computation behind the final answer.
The concrete result is narrower and more useful: among the systems tested, GPT-6 Astra reaches similar solution depths with a much smaller, more direct visible structure, while the same compression makes its traces harder for weaker models to use.