Agentic Table Benchmark Reports

Measured results for the Agentic Table agents. Each agent has its own section: a cover page with the results at a glance and the charts, and the full report below it.

Library Answers

Returns the approved answer from the Answer Library verbatim, or refuses. Best result: Claude Opus 5.5 at low reasoning, right outcome on 97% of 325 questions (October 2026).

• Library Answers Benchmarks

Subagent Answer

Composes an answer from eligible library sources and flags gaps. Claude Opus 5.5 at low reasoning matches the approved answers best on both content (99% of questions the library answers) and tone; GPT-6.1 Sol at low reasoning flags the gap on 72% of questions the library does not answer, while the other models mostly compose an answer (October 2026).

• Subagent Answer Benchmarks

Question Extraction

Turns an uploaded questionnaire document into the list of questions the table answers. On the 11 documents every model completes, the best models capture 93–94% of questions; GPT-6 Sol at its default reasoning captures 93%, the most of the models that complete all 18 standard documents (October 2026).

• Question Extraction Benchmarks