{"id":9104,"date":"2026-11-16T08:00:49","date_gmt":"2026-11-16T08:00:49","guid":{"rendered":"https:\/\/i10xnews.kinsta.cloud\/ai-agents-spreadsheets-gdp-xlsx-study"},"modified":"2026-11-16T08:00:49","modified_gmt":"2026-11-16T08:00:49","slug":"ai-agents-spreadsheets-gdp-xlsx-study","status":"publish","type":"post","link":"https:\/\/i10x.ai\/news\/ai-agents-spreadsheets-gdp-xlsx-study","title":{"rendered":"Study: Best AI Agent Passes Only 38% of Real Spreadsheet Tasks"},"content":{"rendered":"\n<h2 class=\"wp-block-heading\">\u26a1 Quick Take<\/h2>\n\n<ul class=\"wp-block-list\"><li><strong>The finding:<\/strong> Across 20 model configurations, the best, Google&#8217;s Gemini 4 Argon, fully solves only 38.3% of attempts on 70 expert spreadsheet requests in the GDP.xlsx benchmark. The median configuration passes 19.3%, and Mistral Large 3 passes none.<\/li><li><strong>Why it matters:<\/strong> Every task was written by a domain professional around a workflow where frontier models had failed them, and an answer passes only if it meets every rubric criterion, including finding every item of a set and nothing extra.<\/li><li><strong>What it means:<\/strong> Agents that read workbooks still miss evidence that sits on another tab, behind a filter, in a cell comment or in a pasted diagram. AI answers about a spreadsheet are a draft to check, not a final figure.<\/li><\/ul>\n\n<h2 class=\"wp-block-heading\">In plain English<\/h2>\n\n\n<p class=\"wp-block-paragraph\">A lot of business decisions live in spreadsheets. Researchers at Surge AI, the company behind this test, gave 20 AI setups real questions that office professionals had about their own spreadsheets, together with the full files. The best AI gave a fully correct answer in fewer than four out of ten tries, and the typical one in about two out of ten. The AIs often found part of the answer but missed items tucked away on another tab, in a note or in a picture pasted into the file. If you ask an AI about a big spreadsheet, check its answer before you rely on it.<\/p>\n\n\n<h2 class=\"wp-block-heading\">\ud83d\udd2c What the study shows<\/h2>\n\n\n<p class=\"wp-block-paragraph\">A professional who inherits a workbook usually inherits a question with it: which tenancies change rent in the next three years, which products are cleared for production, which form fields capture housebound patients. Suhaas Garre, Sushant Mehta and Edwin Chen of Surge AI turned that situation into a benchmark. GDP.xlsx pairs 70 workplace requests from 13 domains with the complete Excel workbook or CSV file each one refers to, 2.18 million populated cells in total, and grades the answers against 791 expert criteria. The paper, posted to arXiv on November 13, 2026, documents the launch leaderboard of September 30, 2026 and has been accepted to the Agent Evaluation Science Symposium in New York.<\/p>\n\n<ul class=\"wp-block-list\"><li><strong>Every configuration fails most attempts:<\/strong> Gemini 4 Argon (High) leads with 38.3% strict pass@1, with a task-sampling standard error of at most 5.8 points. Claude Opus 5.5 (30.3%) and Claude Sonnet 5.5 (29.1%) follow. No other configuration exceeds 25%, and the median of 19.3% is half the leader&#8217;s score.<\/li><li><strong>Newer versions do better:<\/strong> Claude Sonnet 5.5 beats Claude Sonnet 5 (29.1% versus 16.9%), GPT 6.1 Sol beats GPT 6 Sol (22.0% versus 17.1%), and Gemini 4 Argon scores 28 points above Gemini 3.1 Pro (10.3%).<\/li><li><strong>A long tail:<\/strong> DeepSeek V4 Pro (Max) passes 6.9% of attempts, Nemotron 3 Ultra 4.9% and Mistral Large 3 0.0%. Grok 4.7 (xHigh) sits mid-pack at 22.3%.<\/li><li><strong>The misses are recall failures:<\/strong> In all four failed attempts the authors dissect, the answer left out evidence the file holds. Muse Spark 1.3 listed 6 of 9 required tenancies with perfect precision but skipped the three that qualify only through a tenant break, recorded on a separate tab where a filter hides one of them. GPT 6 Astra named 3 of 8 housebound codes in a clinical data dictionary. Gemini 3.8 Flash picked the wrong equipment cabinet because the right one is labeled only in a diagram pasted onto a worksheet.<\/li><li><strong>The files are messy on purpose:<\/strong> 52 of the 58 workbooks have more than one worksheet, 42 use cell fill colors and 17 contain hidden sheets, rows or columns. In 21 tasks the answer depends on layout or formatting, and 58 of the 70 requests never use the word &#8220;tab&#8221; or &#8220;sheet&#8221;.<\/li><\/ul>\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1344\" height=\"768\" src=\"https:\/\/i10xnews.kinsta.cloud\/wp-content\/uploads\/2026\/11\/chart-gdp-xlsx-pass-rates.webp\" alt=\"Horizontal bar chart of strict pass@1 on GDP.xlsx for 20 model configurations: Gemini 4 Argon 38.3%, Claude Opus 5.5 30.3%, Claude Sonnet 5.5 29.1%, falling to DeepSeek V4 Pro 6.9%, Nemotron 3 Ultra 4.9% and Mistral Large 3 0.0%.\" class=\"wp-image-9103\" srcset=\"https:\/\/i10x.ai\/news\/wp-content\/uploads\/2026\/11\/chart-gdp-xlsx-pass-rates.webp 1344w, https:\/\/i10x.ai\/news\/wp-content\/uploads\/2026\/11\/chart-gdp-xlsx-pass-rates-300x171.webp 300w, https:\/\/i10x.ai\/news\/wp-content\/uploads\/2026\/11\/chart-gdp-xlsx-pass-rates-1024x585.webp 1024w, https:\/\/i10x.ai\/news\/wp-content\/uploads\/2026\/11\/chart-gdp-xlsx-pass-rates-768x439.webp 768w\" sizes=\"auto, (max-width: 1344px) 100vw, 1344px\" \/><figcaption class=\"wp-element-caption\">i10X chart based on data from Garre, Mehta and Chen (2026), arXiv:2611.17766, Figure 4. CC BY 4.0.<\/figcaption><\/figure>\n\n\n<figure class=\"wp-block-table\"><table><thead><tr><th>Model configuration<\/th><th>Strict pass@1<\/th><\/tr><\/thead><tbody><tr><td>Gemini 4 Argon (High)<\/td><td>38.3%<\/td><\/tr><tr><td>Claude Opus 5.5 (Adaptive\/Max)<\/td><td>30.3%<\/td><\/tr><tr><td>Claude Sonnet 5.5 (Adaptive\/Max)<\/td><td>29.1%<\/td><\/tr><tr><td>Claude Fable 5.1 (Adaptive\/Max)<\/td><td>23.1%<\/td><\/tr><tr><td>GPT 6 Astra (Max)<\/td><td>22.9%<\/td><\/tr><tr><td>Grok 4.7 (xHigh)<\/td><td>22.3%<\/td><\/tr><tr><td>GPT 6.1 Sol (Max)<\/td><td>22.0%<\/td><\/tr><tr><td>Gemini 3.1 Pro (High)<\/td><td>10.3%<\/td><\/tr><tr><td>DeepSeek V4 Pro (Max)<\/td><td>6.9%<\/td><\/tr><\/tbody><\/table><figcaption class=\"wp-element-caption\">Selected rows of Figure 4 in Garre, Mehta and Chen (2026), arXiv:2611.17766. Launch leaderboard of September 30, 2026, five attempts per task, 70 tasks.<\/figcaption><\/figure>\n\n\n<h2 class=\"wp-block-heading\">\u2705 Why it matters: strict grading on real files<\/h2>\n\n\n<p class=\"wp-block-paragraph\">Most spreadsheet benchmarks check whether a formula or an edit was produced, or compare a short answer with a reference string. GDP.xlsx grades the way a colleague would. In 39 tasks the rubric contains both a completeness criterion, which requires every member of a set, and an exclusivity criterion, which forbids anything else, so an answer must reach full recall and full precision. Prohibitions name plausible mistakes, such as using GBP amounts in place of USD. An agentic judge can open the original file and run code to verify each claim, but may not complete the task itself or credit content the response does not contain, and the same judge model and instructions grade every attempt of every configuration.<\/p>\n\n\n<p class=\"wp-block-paragraph\">The tasks themselves are hard to game. Professionals wrote them around workflows in which frontier models had failed them, each passed three rounds of review, and the requests and rubrics appeared in no public source before release. Only three tasks are public; the other 67 are held out, with a canary string to keep them out of training data. Each configuration gets five attempts per task, and the paper reports the statistical uncertainty of its scores.<\/p>\n\n\n<h2 class=\"wp-block-heading\">\ud83e\uddea Method in brief<\/h2>\n\n<ul class=\"wp-block-list\"><li><strong>Tasks:<\/strong> 70 requests across 13 domains, led by manufacturing and supply chain (10 tasks), construction, healthcare and human resources (8 each). Inputs are 58 Excel workbooks and 12 CSV files. Rubrics have 3 to 30 binary criteria, about 11 on average.<\/li><li><strong>Harness:<\/strong> An agent loop built on the OpenHands Software Agent SDK, run in a container with a terminal, a file editor, Python with openpyxl and pandas, LibreOffice and Tesseract OCR. No spreadsheet-specific instructions. The container can reach only the model provider&#8217;s API. Limits: 200 turns, 60 minutes and 600 seconds per model call.<\/li><li><strong>Judge:<\/strong> The OpenCode agent, with one judge model and one set of instructions for all attempts. It must write its reasoning before each binary verdict.<\/li><li><strong>Metric:<\/strong> Strict pass@1: the share of attempts that meet every criterion, averaged over tasks. Attempts that end in an error or time-out are excluded; attempts that hit the turn limit count as empty responses.<\/li><li><strong>Models:<\/strong> 20 configurations from 12 developers as of September 2026, including Google, Anthropic, OpenAI, xAI, DeepSeek, Mistral and NVIDIA.<\/li><\/ul>\n\n<h2 class=\"wp-block-heading\">\u2753 What remains open<\/h2>\n\n<ul class=\"wp-block-list\"><li><strong>The judge is not named:<\/strong> The paper says one judge model grades everything but does not say which one, and it reports no agreement check between the judge and human graders.<\/li><li><strong>Close ranks are noisy:<\/strong> With 70 tasks, the standard error for the leader is up to 5.8 points. Configurations one or two points apart, such as the cluster between 20% and 23%, cannot be ranked with confidence.<\/li><li><strong>Limited replication:<\/strong> The harness and three tasks are public under Apache 2.0, but the 67 held-out tasks are not, so outsiders can only rerun a small slice.<\/li><li><strong>One harness, one moment:<\/strong> Results reflect a general agent harness without spreadsheet tools and the models available in September 2026. Structure-aware tools or prompts could change the picture, which the authors themselves suggest testing.<\/li><li><strong>Who built it:<\/strong> The authors work at Surge AI, which also runs the public leaderboard. The paper has no separate funding or competing-interests statement. It is a preprint accepted to a symposium, not a journal paper.<\/li><\/ul>\n\n<h2 class=\"wp-block-heading\">\ud83e\udded Context: beyond formula benchmarks<\/h2>\n\n\n<p class=\"wp-block-paragraph\">Earlier work such as SpreadsheetBench, SheetCopilot and InstructExcel grades whether an instruction was executed, table question answering sets like WikiTableQuestions and TAT-QA score short answers against references, and GDPval and DAYJOB grade end-to-end professional deliverables. GDP.xlsx targets the step before any of that: finding all the evidence in one messy file. It belongs to a family of Surge AI benchmarks that includes GDP.pdf, Chartography, HANDBOOK.md and DAYJOB. For a similar expert-built test on coding, see our report on <a href=\"\/news\/coding-agents-expert-tasks-expertcrowd-swe-study\">ExpertCrowd-SWE<\/a>, and for spreadsheets in finance operations, our piece on <a href=\"\/news\/local-7b-models-enterprise-finance-reconciliation\">local models in enterprise finance workflows<\/a>.<\/p>\n\n\n<h2 class=\"wp-block-heading\">\ud83d\udee0\ufe0f What it means for teams that live in spreadsheets<\/h2>\n\n<ul class=\"wp-block-list\"><li><strong>Business users:<\/strong> Ask the AI to say where each item came from (tab and cell) and whether it checked every tab, comment and filter. Check list-type answers for missing items, not only wrong ones.<\/li><li><strong>Agent builders:<\/strong> Add an explicit retrieval step: list every tab, comment, filter and embedded image before answering, then match the answer against each condition in the request. The authors note that loading a workbook into data frames keeps values but drops formats, comments and drawings.<\/li><li><strong>Buyers:<\/strong> Model choice matters: scores range from 38.3% to 0.0% on the same tasks. <a href=\"\/news\/company\/google\">Google&#8217;s<\/a> <a href=\"\/news\/product\/gemini\">Gemini<\/a> and Anthropic&#8217;s <a href=\"\/news\/product\/claude\">Claude<\/a> models lead this leaderboard, but even the leader fails most attempts.<\/li><\/ul>\n\n<p class=\"wp-block-paragraph\">More on our <a href=\"\/news\/category\/studies\/benchmarks-evals\">Benchmarks &amp; Evals<\/a> studies page.<\/p>\n\n\n<h2 class=\"wp-block-heading\">\u270d\ufe0f About the analysis<\/h2>\n\n\n<p class=\"wp-block-paragraph\">This article is based on the full text of Garre, Mehta and Chen (2026), &#8220;GDP.xlsx: Benchmarking Frontier Models on Professional Workbooks,&#8221; arXiv:2611.17766v1 (cs.CL, submitted November 13, 2026). All numbers come from the paper. The chart was redrawn by i10X from Figure 4 of the paper, which is published under a CC BY 4.0 license. The evaluation harness is at github.com\/surge-ai\/gdp-xlsx and the three public tasks are on Hugging Face (surgeai\/GDP.xlsx). No conflicts of interest stated beyond the authors&#8217; affiliation with Surge AI. AI tools assisted with research and drafting; selection, fact-checking and final editing were done by the i10X editorial team.<\/p>\n\n\n<h2 class=\"wp-block-heading\">\ud83d\udd2d i10x Perspective<\/h2>\n\n\n<p class=\"wp-block-paragraph\">This is a sobering number for anyone selling AI as an analyst. The failures are not wild hallucinations; they are answers that look complete and are not. That is the most dangerous kind of error in a spreadsheet, because nobody double-checks a tidy table. Until agents learn to search a file exhaustively, the human reviewer is still part of the product.<\/p>\n\n","protected":false},"excerpt":{"rendered":"<p>Surge AI&#8217;s GDP.xlsx gives 20 model setups 70 expert requests over full workbooks. Gemini 4 Argon leads at 38.3% strict pass@1, and the median setup passes only 19.3%.<\/p>\n","protected":false},"author":2,"featured_media":9102,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"paper_title":"GDP.xlsx: Benchmarking Frontier Models on Professional Workbooks","paper_url":"https:\/\/arxiv.org\/abs\/2611.17766","pdf_url":"https:\/\/arxiv.org\/pdf\/2611.17766","doi":"","arxiv_id":"2611.17766","authors":"Suhaas Garre, Sushant Mehta, Edwin Chen","institutions":"Surge AI","paper_date":"2026-11-13","venue":"arXiv (cs.CL), version 1; accepted to the Agent Evaluation Science Symposium (New York City, November 2026)","peer_reviewed":false,"code_url":"https:\/\/github.com\/surge-ai\/gdp-xlsx","data_url":"https:\/\/huggingface.co\/datasets\/surgeai\/GDP.xlsx","funding_coi":"No funding or competing-interests statement in the paper. All authors are at Surge AI, which publishes the benchmark and leaderboard.","figure_license":"Own chart redrawn from Figure 4 (no original figures reused)","paper_license":"CC BY 4.0","replication":"partial","study_models":"20 configurations from 12 developers, including Gemini 4 Argon, Gemini 3.8 Flash, Gemini 3.1 Pro, Claude Opus 5.5, Claude Sonnet 5.5, Claude Sonnet 5, Claude Fable 5.1, GPT 6 Astra, GPT 6.1 Sol, GPT 6 Sol, GPT 6 Luna, Grok 4.7, Muse Spark 1.3, GLM 5.3, Qwen 3.8 Max, Hy4 Preview, Kimi K3, DeepSeek V4 Pro, Nemotron 3 Ultra, Mistral Large 3","study_sample":"70 expert-written tasks in 13 domains (58 Excel workbooks, 12 CSV files, 2.18 million populated cells, 791 rubric criteria); five attempts per task; 3 tasks public, 67 held out","footnotes":"","_yoast_wpseo_primary_category":"52","_nx_sec_severity":"","_nx_sec_affected":"","_nx_sec_status":"","_nx_sec_kicker":"","_nx_sec_type":"","_nx_sec_vendor":"","_nx_sec_fixed":"","_nx_sec_action":""},"categories":[53,52],"tags":[],"companies":[39,46,37,38,40],"products":[51,49,48],"class_list":["post-9104","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-benchmarks-evals","category-studies","company-anthropic","company-deepseek","company-google","company-open-ai","company-x-ai","product-claude","product-gemini","product-grok"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v28.7 - https:\/\/yoast.com\/product\/yoast-seo-wordpress\/ -->\n<title>Best AI Agent Passes Only 38% of Spreadsheet Tasks - i10X News<\/title>\n<meta name=\"description\" content=\"Surge AI&#039;s GDP.xlsx tests 20 AI setups on 70 expert workbook requests. Gemini 4 Argon leads at 38.3% strict pass@1, and the median setup passes just 19.3%.\" \/>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/i10xnews.kinsta.cloud\/ai-agents-spreadsheets-gdp-xlsx-study\" \/>\n<meta property=\"og:locale\" content=\"en_US\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"Best AI Agent Passes Only 38% of Spreadsheet Tasks - i10X News\" \/>\n<meta property=\"og:description\" content=\"Surge AI&#039;s GDP.xlsx tests 20 AI setups on 70 expert workbook requests. Gemini 4 Argon leads at 38.3% strict pass@1, and the median setup passes just 19.3%.\" \/>\n<meta property=\"og:url\" content=\"https:\/\/i10xnews.kinsta.cloud\/ai-agents-spreadsheets-gdp-xlsx-study\" \/>\n<meta property=\"og:site_name\" content=\"i10X News\" \/>\n<meta property=\"article:published_time\" content=\"2026-11-16T08:00:49+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/i10xnews.kinsta.cloud\/wp-content\/uploads\/2026\/11\/ai-agents-spreadsheets-gdp-xlsx-study-title.webp\" \/>\n\t<meta property=\"og:image:width\" content=\"1344\" \/>\n\t<meta property=\"og:image:height\" content=\"768\" \/>\n\t<meta property=\"og:image:type\" content=\"image\/webp\" \/>\n<meta name=\"author\" content=\"Christopher Ort\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:image\" content=\"https:\/\/i10x.ai\/news\/wp-content\/uploads\/2026\/10\/i10x-news-og-default-1200x630-1.png\" \/>\n<meta name=\"twitter:creator\" content=\"@i10x_ai\" \/>\n<meta name=\"twitter:site\" content=\"@i10x_ai\" \/>\n<meta name=\"twitter:label1\" content=\"Written by\" \/>\n\t<meta name=\"twitter:data1\" content=\"Christopher Ort\" \/>\n\t<meta name=\"twitter:label2\" content=\"Est. reading time\" \/>\n\t<meta name=\"twitter:data2\" content=\"8 minutes\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\\\/\\\/schema.org\",\"@graph\":[{\"@type\":\"NewsArticle\",\"@id\":\"https:\\\/\\\/i10xnews.kinsta.cloud\\\/ai-agents-spreadsheets-gdp-xlsx-study#article\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/i10xnews.kinsta.cloud\\\/ai-agents-spreadsheets-gdp-xlsx-study\"},\"author\":{\"name\":\"Christopher Ort\",\"@id\":\"https:\\\/\\\/i10x.ai\\\/news\\\/#\\\/schema\\\/person\\\/d410bb7ee72febefca2c9c369939219a\"},\"headline\":\"Study: Best AI Agent Passes Only 38% of Real Spreadsheet Tasks\",\"datePublished\":\"2026-11-16T08:00:49+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\\\/\\\/i10xnews.kinsta.cloud\\\/ai-agents-spreadsheets-gdp-xlsx-study\"},\"wordCount\":1553,\"publisher\":{\"@id\":\"https:\\\/\\\/i10x.ai\\\/news\\\/#organization\"},\"image\":{\"@id\":\"https:\\\/\\\/i10xnews.kinsta.cloud\\\/ai-agents-spreadsheets-gdp-xlsx-study#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/i10x.ai\\\/news\\\/wp-content\\\/uploads\\\/2026\\\/11\\\/ai-agents-spreadsheets-gdp-xlsx-study-title.webp\",\"articleSection\":[\"Benchmarks &amp; Evals\",\"Studies\"],\"inLanguage\":\"en-US\",\"dateModified\":\"2026-11-16T08:00:49+00:00\"},{\"@type\":\"WebPage\",\"@id\":\"https:\\\/\\\/i10xnews.kinsta.cloud\\\/ai-agents-spreadsheets-gdp-xlsx-study\",\"url\":\"https:\\\/\\\/i10xnews.kinsta.cloud\\\/ai-agents-spreadsheets-gdp-xlsx-study\",\"name\":\"Best AI Agent Passes Only 38% of Spreadsheet Tasks - i10X News\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/i10x.ai\\\/news\\\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\\\/\\\/i10xnews.kinsta.cloud\\\/ai-agents-spreadsheets-gdp-xlsx-study#primaryimage\"},\"image\":{\"@id\":\"https:\\\/\\\/i10xnews.kinsta.cloud\\\/ai-agents-spreadsheets-gdp-xlsx-study#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/i10x.ai\\\/news\\\/wp-content\\\/uploads\\\/2026\\\/11\\\/ai-agents-spreadsheets-gdp-xlsx-study-title.webp\",\"datePublished\":\"2026-11-16T08:00:49+00:00\",\"description\":\"Surge AI's GDP.xlsx tests 20 AI setups on 70 expert workbook requests. Gemini 4 Argon leads at 38.3% strict pass@1, and the median setup passes just 19.3%.\",\"breadcrumb\":{\"@id\":\"https:\\\/\\\/i10xnews.kinsta.cloud\\\/ai-agents-spreadsheets-gdp-xlsx-study#breadcrumb\"},\"inLanguage\":\"en-US\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\\\/\\\/i10xnews.kinsta.cloud\\\/ai-agents-spreadsheets-gdp-xlsx-study\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"en-US\",\"@id\":\"https:\\\/\\\/i10xnews.kinsta.cloud\\\/ai-agents-spreadsheets-gdp-xlsx-study#primaryimage\",\"url\":\"https:\\\/\\\/i10x.ai\\\/news\\\/wp-content\\\/uploads\\\/2026\\\/11\\\/ai-agents-spreadsheets-gdp-xlsx-study-title.webp\",\"contentUrl\":\"https:\\\/\\\/i10x.ai\\\/news\\\/wp-content\\\/uploads\\\/2026\\\/11\\\/ai-agents-spreadsheets-gdp-xlsx-study-title.webp\",\"width\":1344,\"height\":768,\"caption\":\"Image: AI-generated, i10X\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\\\/\\\/i10xnews.kinsta.cloud\\\/ai-agents-spreadsheets-gdp-xlsx-study#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"Home\",\"item\":\"https:\\\/\\\/i10x.ai\\\/news\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"Study: Best AI Agent Passes Only 38% of Real Spreadsheet Tasks\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\\\/\\\/i10x.ai\\\/news\\\/#website\",\"url\":\"https:\\\/\\\/i10x.ai\\\/news\",\"name\":\"i10X News\",\"description\":\"Stay updated with the latest news and announcements from i10X.\",\"publisher\":{\"@id\":\"https:\\\/\\\/i10x.ai\\\/news\\\/#organization\"},\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\\\/\\\/i10x.ai\\\/news?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"en-US\"},{\"@type\":\"Organization\",\"@id\":\"https:\\\/\\\/i10x.ai\\\/news\\\/#organization\",\"name\":\"i10X\",\"url\":\"https:\\\/\\\/i10x.ai\\\/news\\\/\",\"logo\":{\"@type\":\"ImageObject\",\"inLanguage\":\"en-US\",\"@id\":\"https:\\\/\\\/i10x.ai\\\/news\\\/#\\\/schema\\\/logo\\\/image\\\/\",\"url\":\"https:\\\/\\\/i10xnews.kinsta.cloud\\\/wp-content\\\/uploads\\\/2026\\\/10\\\/i10x-logo-180.png\",\"contentUrl\":\"https:\\\/\\\/i10xnews.kinsta.cloud\\\/wp-content\\\/uploads\\\/2026\\\/10\\\/i10x-logo-180.png\",\"width\":180,\"height\":180,\"caption\":\"i10X\"},\"image\":{\"@id\":\"https:\\\/\\\/i10x.ai\\\/news\\\/#\\\/schema\\\/logo\\\/image\\\/\"},\"sameAs\":[\"https:\\\/\\\/x.com\\\/i10x_ai\",\"https:\\\/\\\/www.linkedin.com\\\/company\\\/i10x-ai\",\"https:\\\/\\\/instagram.com\\\/i10x.ai\",\"https:\\\/\\\/www.youtube.com\\\/@i10X_ai_official\"]},{\"@type\":\"Person\",\"@id\":\"https:\\\/\\\/i10x.ai\\\/news\\\/#\\\/schema\\\/person\\\/d410bb7ee72febefca2c9c369939219a\",\"name\":\"Christopher Ort\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"en-US\",\"@id\":\"https:\\\/\\\/i10x.ai\\\/news\\\/wp-content\\\/uploads\\\/2026\\\/10\\\/christopher-ort-192.webp\",\"url\":\"https:\\\/\\\/i10x.ai\\\/news\\\/wp-content\\\/uploads\\\/2026\\\/10\\\/christopher-ort-192.webp\",\"contentUrl\":\"https:\\\/\\\/i10x.ai\\\/news\\\/wp-content\\\/uploads\\\/2026\\\/10\\\/christopher-ort-192.webp\",\"caption\":\"Christopher Ort\"},\"description\":\"Christopher Ort handles SEO, AI agent workflows and PR & partnerships at i10x.ai, the all-in-one AI workspace built by i10X Global Pte. Ltd. in Singapore. He is a graduate of Friedrich-Alexander-Universit\u00e4t Erlangen-N\u00fcrnberg (FAU). He is the author of \u201cThe Style Penalty\u201d, a study published on SSRN that analyses 1,576 evaluation decisions by four frontier AI models to show how AI-written CVs fare in automated hiring screens; the research has been covered by heise online and t3n.\",\"url\":\"https:\\\/\\\/i10x.ai\\\/news\\\/author\\\/christopher-ort\",\"sameAs\":[\"https:\\\/\\\/i10x.ai\\\/blog\\\/author\\\/christopher-ort\"],\"worksFor\":{\"@id\":\"https:\\\/\\\/i10x.ai\\\/news\\\/#organization\"},\"alumniOf\":{\"@type\":\"EducationalOrganization\",\"name\":\"Friedrich-Alexander-Universit\u00e4t Erlangen-N\u00fcrnberg (FAU)\",\"url\":\"https:\\\/\\\/www.fau.eu\"},\"subjectOf\":[{\"@type\":\"CreativeWork\",\"url\":\"https:\\\/\\\/papers.ssrn.com\\\/abstract=7426699\",\"name\":\"The Style Penalty: How AI Resume-Writing Tools Shape Outcomes in Automated Hiring Screens\"},{\"@type\":\"CreativeWork\",\"url\":\"https:\\\/\\\/www.heise.de\\\/news\\\/KI-Chatbots-bewerten-KI-Lebenslaeufe-Claude-findet-sich-besser-als-GPT-11317531.html\",\"name\":\"KI-Chatbots bewerten KI-Lebensl\u00e4ufe: Claude findet sich besser als GPT\"},{\"@type\":\"CreativeWork\",\"url\":\"https:\\\/\\\/t3n.de\\\/news\\\/bewerbung-schreiben-ki-5-fehler-1747190\\\/\",\"name\":\"Bewerbung mit KI schreiben: Diese 5 Fehler solltet ihr laut Studien vermeiden\"}]}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"Best AI Agent Passes Only 38% of Spreadsheet Tasks - i10X News","description":"Surge AI's GDP.xlsx tests 20 AI setups on 70 expert workbook requests. Gemini 4 Argon leads at 38.3% strict pass@1, and the median setup passes just 19.3%.","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/i10xnews.kinsta.cloud\/ai-agents-spreadsheets-gdp-xlsx-study","og_locale":"en_US","og_type":"article","og_title":"Best AI Agent Passes Only 38% of Spreadsheet Tasks - i10X News","og_description":"Surge AI's GDP.xlsx tests 20 AI setups on 70 expert workbook requests. Gemini 4 Argon leads at 38.3% strict pass@1, and the median setup passes just 19.3%.","og_url":"https:\/\/i10xnews.kinsta.cloud\/ai-agents-spreadsheets-gdp-xlsx-study","og_site_name":"i10X News","article_published_time":"2026-11-16T08:00:49+00:00","og_image":[{"width":1344,"height":768,"url":"https:\/\/i10xnews.kinsta.cloud\/wp-content\/uploads\/2026\/11\/ai-agents-spreadsheets-gdp-xlsx-study-title.webp","type":"image\/webp"}],"author":"Christopher Ort","twitter_card":"summary_large_image","twitter_image":"https:\/\/i10x.ai\/news\/wp-content\/uploads\/2026\/10\/i10x-news-og-default-1200x630-1.png","twitter_creator":"@i10x_ai","twitter_site":"@i10x_ai","twitter_misc":{"Written by":"Christopher Ort","Est. reading time":"8 minutes"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"NewsArticle","@id":"https:\/\/i10xnews.kinsta.cloud\/ai-agents-spreadsheets-gdp-xlsx-study#article","isPartOf":{"@id":"https:\/\/i10xnews.kinsta.cloud\/ai-agents-spreadsheets-gdp-xlsx-study"},"author":{"name":"Christopher Ort","@id":"https:\/\/i10x.ai\/news\/#\/schema\/person\/d410bb7ee72febefca2c9c369939219a"},"headline":"Study: Best AI Agent Passes Only 38% of Real Spreadsheet Tasks","datePublished":"2026-11-16T08:00:49+00:00","mainEntityOfPage":{"@id":"https:\/\/i10xnews.kinsta.cloud\/ai-agents-spreadsheets-gdp-xlsx-study"},"wordCount":1553,"publisher":{"@id":"https:\/\/i10x.ai\/news\/#organization"},"image":{"@id":"https:\/\/i10xnews.kinsta.cloud\/ai-agents-spreadsheets-gdp-xlsx-study#primaryimage"},"thumbnailUrl":"https:\/\/i10x.ai\/news\/wp-content\/uploads\/2026\/11\/ai-agents-spreadsheets-gdp-xlsx-study-title.webp","articleSection":["Benchmarks &amp; Evals","Studies"],"inLanguage":"en-US","dateModified":"2026-11-16T08:00:49+00:00"},{"@type":"WebPage","@id":"https:\/\/i10xnews.kinsta.cloud\/ai-agents-spreadsheets-gdp-xlsx-study","url":"https:\/\/i10xnews.kinsta.cloud\/ai-agents-spreadsheets-gdp-xlsx-study","name":"Best AI Agent Passes Only 38% of Spreadsheet Tasks - i10X News","isPartOf":{"@id":"https:\/\/i10x.ai\/news\/#website"},"primaryImageOfPage":{"@id":"https:\/\/i10xnews.kinsta.cloud\/ai-agents-spreadsheets-gdp-xlsx-study#primaryimage"},"image":{"@id":"https:\/\/i10xnews.kinsta.cloud\/ai-agents-spreadsheets-gdp-xlsx-study#primaryimage"},"thumbnailUrl":"https:\/\/i10x.ai\/news\/wp-content\/uploads\/2026\/11\/ai-agents-spreadsheets-gdp-xlsx-study-title.webp","datePublished":"2026-11-16T08:00:49+00:00","description":"Surge AI's GDP.xlsx tests 20 AI setups on 70 expert workbook requests. Gemini 4 Argon leads at 38.3% strict pass@1, and the median setup passes just 19.3%.","breadcrumb":{"@id":"https:\/\/i10xnews.kinsta.cloud\/ai-agents-spreadsheets-gdp-xlsx-study#breadcrumb"},"inLanguage":"en-US","potentialAction":[{"@type":"ReadAction","target":["https:\/\/i10xnews.kinsta.cloud\/ai-agents-spreadsheets-gdp-xlsx-study"]}]},{"@type":"ImageObject","inLanguage":"en-US","@id":"https:\/\/i10xnews.kinsta.cloud\/ai-agents-spreadsheets-gdp-xlsx-study#primaryimage","url":"https:\/\/i10x.ai\/news\/wp-content\/uploads\/2026\/11\/ai-agents-spreadsheets-gdp-xlsx-study-title.webp","contentUrl":"https:\/\/i10x.ai\/news\/wp-content\/uploads\/2026\/11\/ai-agents-spreadsheets-gdp-xlsx-study-title.webp","width":1344,"height":768,"caption":"Image: AI-generated, i10X"},{"@type":"BreadcrumbList","@id":"https:\/\/i10xnews.kinsta.cloud\/ai-agents-spreadsheets-gdp-xlsx-study#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https:\/\/i10x.ai\/news"},{"@type":"ListItem","position":2,"name":"Study: Best AI Agent Passes Only 38% of Real Spreadsheet Tasks"}]},{"@type":"WebSite","@id":"https:\/\/i10x.ai\/news\/#website","url":"https:\/\/i10x.ai\/news","name":"i10X News","description":"Stay updated with the latest news and announcements from i10X.","publisher":{"@id":"https:\/\/i10x.ai\/news\/#organization"},"potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/i10x.ai\/news?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"en-US"},{"@type":"Organization","@id":"https:\/\/i10x.ai\/news\/#organization","name":"i10X","url":"https:\/\/i10x.ai\/news\/","logo":{"@type":"ImageObject","inLanguage":"en-US","@id":"https:\/\/i10x.ai\/news\/#\/schema\/logo\/image\/","url":"https:\/\/i10xnews.kinsta.cloud\/wp-content\/uploads\/2026\/10\/i10x-logo-180.png","contentUrl":"https:\/\/i10xnews.kinsta.cloud\/wp-content\/uploads\/2026\/10\/i10x-logo-180.png","width":180,"height":180,"caption":"i10X"},"image":{"@id":"https:\/\/i10x.ai\/news\/#\/schema\/logo\/image\/"},"sameAs":["https:\/\/x.com\/i10x_ai","https:\/\/www.linkedin.com\/company\/i10x-ai","https:\/\/instagram.com\/i10x.ai","https:\/\/www.youtube.com\/@i10X_ai_official"]},{"@type":"Person","@id":"https:\/\/i10x.ai\/news\/#\/schema\/person\/d410bb7ee72febefca2c9c369939219a","name":"Christopher Ort","image":{"@type":"ImageObject","inLanguage":"en-US","@id":"https:\/\/i10x.ai\/news\/wp-content\/uploads\/2026\/10\/christopher-ort-192.webp","url":"https:\/\/i10x.ai\/news\/wp-content\/uploads\/2026\/10\/christopher-ort-192.webp","contentUrl":"https:\/\/i10x.ai\/news\/wp-content\/uploads\/2026\/10\/christopher-ort-192.webp","caption":"Christopher Ort"},"description":"Christopher Ort handles SEO, AI agent workflows and PR & partnerships at i10x.ai, the all-in-one AI workspace built by i10X Global Pte. Ltd. in Singapore. He is a graduate of Friedrich-Alexander-Universit\u00e4t Erlangen-N\u00fcrnberg (FAU). He is the author of \u201cThe Style Penalty\u201d, a study published on SSRN that analyses 1,576 evaluation decisions by four frontier AI models to show how AI-written CVs fare in automated hiring screens; the research has been covered by heise online and t3n.","url":"https:\/\/i10x.ai\/news\/author\/christopher-ort","sameAs":["https:\/\/i10x.ai\/blog\/author\/christopher-ort"],"worksFor":{"@id":"https:\/\/i10x.ai\/news\/#organization"},"alumniOf":{"@type":"EducationalOrganization","name":"Friedrich-Alexander-Universit\u00e4t Erlangen-N\u00fcrnberg (FAU)","url":"https:\/\/www.fau.eu"},"subjectOf":[{"@type":"CreativeWork","url":"https:\/\/papers.ssrn.com\/abstract=7426699","name":"The Style Penalty: How AI Resume-Writing Tools Shape Outcomes in Automated Hiring Screens"},{"@type":"CreativeWork","url":"https:\/\/www.heise.de\/news\/KI-Chatbots-bewerten-KI-Lebenslaeufe-Claude-findet-sich-besser-als-GPT-11317531.html","name":"KI-Chatbots bewerten KI-Lebensl\u00e4ufe: Claude findet sich besser als GPT"},{"@type":"CreativeWork","url":"https:\/\/t3n.de\/news\/bewerbung-schreiben-ki-5-fehler-1747190\/","name":"Bewerbung mit KI schreiben: Diese 5 Fehler solltet ihr laut Studien vermeiden"}]}]}},"_links":{"self":[{"href":"https:\/\/i10x.ai\/news\/wp-json\/wp\/v2\/posts\/9104","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/i10x.ai\/news\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/i10x.ai\/news\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/i10x.ai\/news\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/i10x.ai\/news\/wp-json\/wp\/v2\/comments?post=9104"}],"version-history":[{"count":0,"href":"https:\/\/i10x.ai\/news\/wp-json\/wp\/v2\/posts\/9104\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/i10x.ai\/news\/wp-json\/wp\/v2\/media\/9102"}],"wp:attachment":[{"href":"https:\/\/i10x.ai\/news\/wp-json\/wp\/v2\/media?parent=9104"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/i10x.ai\/news\/wp-json\/wp\/v2\/categories?post=9104"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/i10x.ai\/news\/wp-json\/wp\/v2\/tags?post=9104"},{"taxonomy":"company","embeddable":true,"href":"https:\/\/i10x.ai\/news\/wp-json\/wp\/v2\/companies?post=9104"},{"taxonomy":"product","embeddable":true,"href":"https:\/\/i10x.ai\/news\/wp-json\/wp\/v2\/products?post=9104"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}