SecondSourceJudgment rebuilt from primary sources

According to the paper's abstract, Surge AI's new GDP.xlsx benchmark finds the best of 20 model configurations met every grading criterion on one attempt in only 38.3% of 70 real workplace spreadsheet tasks.

Concept illustration: a white cabinet with two drawers has a framed display compartment on top, headed by a black banner with a purple-and-orange S-shaped mark and the lettering SURGE AI on the left and GDP.XLSX on the right. On the shelf below, from left to right, stand a black laptop whose screen reads GEMINI 4 ARGON, a small white table card with a multicoloured G and the lettering GOOGLE, a tall stack of beige sheets with a grid on top and a black bug sitting on it, and a lower stack of brick-red sheets with a grid on top. Printed on the image: 38.3% Spreadsheet tasks fully passed first try: top model vs 19.3% median of 20 configurations (Surge AI).

Full completion of spreadsheet work on a first attempt remains rare on this measure, though Surge AI wrote the tasks and AI judges did the grading. Surge AI is an AI evaluation company. The abstract doesn't give the agreement rate between the AI judges and human experts, or how many runs each task got. The top score went to Google's Gemini 4 Argon, and the median across the 20 configurations was 19.3%. The share counts tasks that passed every criterion on one attempt; it doesn't mean 38.3% of the content was right. Professionals in 13 fields wrote the tasks, and AI judges that can open the files graded them against rubrics written by experts. Only 3 of the 70 tasks are public. The abstract doesn't list how Anthropic's and OpenAI's models scored (arXiv 2611.17766, 2026-11-13). How we checked →

Source and checks

This item ran in the brief of Nov 16, 2026, alongside 10 other items. How we checked it, the quote checks and what was still missing that day are on that issue's How we checked page.

Cite this item

SecondSource (2026). "According to the paper's abstract, Surge AI's new GDP.xlsx benchmark finds the best of 20 model configurations met every grading criterion on one attempt in only 38.3% of 70 real workplace spreadsheet tasks.." https://secondsource.io/en/research/2026-11-16-1/. Accessed access date.