GPT-6 Astra completed 14 of 20 bare Lean 4 theorem tasks in TCSExam-Full, a 100-task benchmark, according to the TCSExam paper. It had no accepted complete resolutions among the suite’s 30 open conjectures. Other agents did submit accepted conjecture results, which the paper’s authors say adapted previously published OpenAI work.
What the 14-of-20 result measures
Each theorem task required a complete proof in Lean 4, a programming language and proof assistant that checks mathematical arguments against formal rules. TCSExam counted a proof only when it passed the formal checks; incomplete submissions received zero on the bare theorem tasks. Astra completed 14 of the 20 tasks, or 70.0%.
The benchmark’s Full suite also included 30 solver-engineering problems, 20 construction tasks and 30 open conjectures. The theorem result is specific to its 20-task set and the tested agent setup. It is not a general accuracy rate for Lean mathematics.
Astra’s 70.0% is its complete-proof result on 20 Lean tasks; 71.9 is its combined score across four different tracks.
The setup behind Astra’s result
TCSExam evaluated GPT-6 Astra with the Codex 0.154.0 harness at maximum effort, allowing one attempt per task and up to 12 hours of development for each attempt. General Internet access was disabled for the theorem and solver tasks; it was allowed for construction and open-conjecture tasks.
A harness is the software setup that connects a model to a task and shapes how it can work on it. The reported result therefore describes Astra running with that harness, effort setting, attempt limit and task budget, rather than a model-only score independent of tools and conditions.
What the 71.9 Full Index combines
The Full Index combines the means from theorem, solver, construction and open-conjecture tracks, weighted 30%, 30%, 30% and 10%, respectively. Astra’s four track results were:
| TCSExam-Full track | Astra result | What it measures |
| Theorem proving | 70.0% (14/20) | Complete accepted Lean 4 proofs |
| Solver engineering | 64.7% | Relative solver score |
| Constructions | 104.9% | Construction-task reward, which can exceed 100% when a reference is surpassed |
| Open conjectures | 0/30 | Accepted complete resolutions |
The Full Index of 71.9 is the weighted aggregate, not Astra’s Lean score. Its components measure different kinds of work, and construction rewards can exceed 100%; the aggregate and the theorem success rate answer different questions.
Other agents had accepted conjecture submissions
Astra had no accepted complete resolutions across the 30 open-conjecture tasks. The paper also records accepted submissions from other agents: Kimi K3.1 Pro had a Ryser submission accepted in 1 of 3 attempts, while DeepSeek V4.1 Flash 1017 had Ryser submissions accepted in 2 of 3 attempts and MIRUP submissions in 3 of 3. The authors say these submissions adapted previously published OpenAI proofs or counterexamples.
Those results make Astra’s zero a result for Astra, not for every agent evaluated on the conjectures.
Lite scores use a different task set
TCSExam-Lite contains 20 tasks, five per track. On Lite, Claude Opus 5.5 scored 72.3 and GPT-6 Astra scored 67.0. Claude Opus 5.5 was evaluated only on Lite, so that comparison does not rank the two models on TCSExam-Full.
TCSExam’s formal theorem benchmark measures a different task family from ARC-AGI-3. NeoTeo’s earlier coverage of Astra’s ARC-AGI-3 scores discussed that separate benchmark.
