{"@context":"https://neupai.io/schema/v0.2","@type":"StructuredNewsArticle","identity":{"article_id":"tech42_20261014_ai-accounting-benchmark","canonical_url":"https://www.tech42.co.kr/%ea%b3%b5%ea%b0%9c%eb%a5%bc-%eb%a7%9d%ec%84%a4%ec%98%80%eb%8b%a4ai-%ec%8a%a4%ed%83%80%ed%8a%b8%ec%97%85%ec%9d%b4-%eb%82%b4%eb%86%93%ec%9d%80-%ed%9a%8c%ea%b3%84%ec%82%ac-vs-ai-%ec%8b%a4/?utm_source=rss&utm_medium=rss&utm_campaign=%25ea%25b3%25b5%25ea%25b0%259c%25eb%25a5%25bc-%25eb%25a7%259d%25ec%2584%25a4%25ec%2598%2580%25eb%258b%25a4ai-%25ec%258a%25a4%25ed%2583%2580%25ed%258a%25b8%25ec%2597%2585%25ec%259d%25b4-%25eb%2582%25b4%25eb%2586%2593%25ec%259d%2580-%25ed%259a%258c%25ea%25b3%2584%25ec%2582%25ac-vs-ai-%25ec%258b%25a4","ai_url":null,"publisher":{"name":"테크42","domain":"www.tech42.co.kr","type":"online"},"author":"정재엽 기자","published_at":"2026-10-14T00:25:28.000Z","updated_at":null,"language":"en","article_type":"straight_news","originality":"self_produced"},"content":{"headline":"“We Hesitated to Publish”…AI Startup's Results on Accountants vs AI Experiment","summary":"In an accounting task benchmark conducted by U.S. AI startup Mercor, AI models were found to be faster, more accurate, and far cheaper than certified public accountants. The researchers suggested that as AI surpasses humans in standardized tasks, productivity measurement standards need to change.","topics":["ai","accounting","productivity","startups"],"geography":["US","KR"],"entities":[{"name":"Mercor","canonical_id":"corp:us:mercor","type":"company","role_in_article":"primary_subject","metadata":{"ticker":null,"parent":null}},{"name":"Ramp","canonical_id":"corp:us:ramp","type":"company","role_in_article":"mentioned","metadata":{"ticker":null,"parent":null}},{"name":"Anthropic","canonical_id":"corp:us:anthropic","type":"company","role_in_article":"mentioned","metadata":{"ticker":null,"parent":null}},{"name":"Claude Opus 5","canonical_id":"product:us:claude-opus-5","type":"product","role_in_article":"mentioned","metadata":{"ticker":null,"parent":"corp:us:anthropic"}},{"name":"OpenAI","canonical_id":"corp:us:openai","type":"company","role_in_article":"mentioned","metadata":{"ticker":null,"parent":null}},{"name":"GPT-5","canonical_id":"product:us:gpt-5","type":"product","role_in_article":"mentioned","metadata":{"ticker":null,"parent":"corp:us:openai"}},{"name":"Claude Fable 5","canonical_id":"product:us:claude-fable-5","type":"product","role_in_article":"mentioned","metadata":{"ticker":null,"parent":"corp:us:anthropic"}}],"claims":[{"id":"c1","statement":"The research team recruited 12 junior accountants holding U.S. Certified Public Accountant (CPA) qualifications","as_of":"2026-10","as_of_explicit":false,"as_of_raw":"October 2026","source_type":"research_paper","comparison":null,"type":"fact","figures":{"value":12,"unit":"명","approximate":false,"converted":null},"expiry_hint":null,"insight":null},{"id":"c2","statement":"The average experience of participants was about 5 years and 6 months","as_of":"2026-10","as_of_explicit":false,"as_of_raw":"October 2026","source_type":"research_paper","comparison":null,"type":"fact","figures":{"value":5.5,"unit":"년","approximate":true,"converted":null},"expiry_hint":null,"insight":null},{"id":"c3","statement":"The tasks were simplified from 'APEX-Accounting,' an accounting benchmark Mercor released jointly with fintech company Ramp in July","as_of":"2026-07","as_of_explicit":true,"as_of_raw":"last July","source_type":"research_paper","comparison":null,"type":"fact","figures":null,"expiry_hint":null,"insight":null},{"id":"c4","statement":"The tasks consisted of four month-end closing scenarios modeled on actual practice, requiring participants to search company work files for the correct figures, perform calculations, and produce a results table","as_of":"2026-10","as_of_explicit":false,"as_of_raw":"October 2026","source_type":"research_paper","comparison":null,"type":"fact","figures":{"value":4,"unit":"개","approximate":false,"converted":null},"expiry_hint":null,"insight":null},{"id":"c5","statement":"Accountants' accuracy varied widely, from 0% to about 90%, with an average of about 37%","as_of":"2026-10","as_of_explicit":false,"as_of_raw":"October 2026","source_type":"research_paper","comparison":null,"type":"fact","figures":{"value":37,"unit":"%","approximate":true,"converted":null},"expiry_hint":null,"insight":null},{"id":"c6","statement":"Most tasks took between 30 and 180 minutes each to complete","as_of":"2026-10","as_of_explicit":false,"as_of_raw":"October 2026","source_type":"research_paper","comparison":null,"type":"fact","figures":{"value":30,"unit":"분","approximate":false,"converted":null},"expiry_hint":null,"insight":null},{"id":"c7","statement":"Anthropic's Claude Opus 5 achieved 100% across all 20 attempts, completing the task within 10 minutes every time","as_of":"2026-10","as_of_explicit":false,"as_of_raw":"October 2026","source_type":"research_paper","comparison":null,"type":"fact","figures":{"value":100,"unit":"%","approximate":false,"converted":null},"expiry_hint":null,"insight":null},{"id":"c8","statement":"The cost to meet one evaluation criterion was about ₩280 (US$0.21) for Claude Opus 5","as_of":"2026-10","as_of_explicit":false,"as_of_raw":"October 2026","source_type":"research_paper","comparison":null,"type":"fact","figures":{"value":280,"unit":"KRW","approximate":true,"converted":{"value":0.21,"unit":"USD"}},"expiry_hint":null,"insight":null},{"id":"c9","statement":"For accountants, converted using the median U.S. accountant wage, the cost was about ₩13,900 (US$10.35), a 49-fold difference","as_of":"2026-10","as_of_explicit":false,"as_of_raw":"October 2026","source_type":"research_paper","comparison":"ai_cost","type":"fact","figures":{"value":13900,"unit":"KRW","approximate":true,"converted":{"value":10.35,"unit":"USD"}},"expiry_hint":null,"insight":null},{"id":"c10","statement":"As recently as 18 months ago, even the top-performing models fell short of the accountants' average score","as_of":"2025-04","as_of_explicit":false,"as_of_raw":"18 months ago","source_type":"research_paper","comparison":"human_average","type":"fact","figures":null,"expiry_hint":null,"insight":null},{"id":"c11","statement":"In spring 2025, OpenAI's reasoning model o3 was the first to surpass that line, and GPT-5 rose to about 69%","as_of":"2025","as_of_explicit":true,"as_of_raw":"spring 2025","source_type":"research_paper","comparison":"human_average","type":"fact","figures":{"value":69,"unit":"%","approximate":true,"converted":null},"expiry_hint":null,"insight":null},{"id":"c12","statement":"Across the full 160-task APEX-Accounting benchmark, the top-ranked model, Claude Fable 5, scored 56.4%","as_of":"2026-10","as_of_explicit":false,"as_of_raw":"October 2026","source_type":"research_paper","comparison":null,"type":"fact","figures":{"value":56.4,"unit":"%","approximate":false,"converted":null},"expiry_hint":null,"insight":null},{"id":"c13","statement":"The proportion of tasks solved correctly in all 8 repeated runs was only 2.6% even for the most consistent model","as_of":"2026-10","as_of_explicit":false,"as_of_raw":"October 2026","source_type":"research_paper","comparison":null,"type":"fact","figures":{"value":2.6,"unit":"%","approximate":false,"converted":null},"expiry_hint":null,"insight":null},{"id":"c14","statement":"The test creators had expected junior accountants to average around 30% and mid-career accountants around 55%","as_of":"2026-10","as_of_explicit":false,"as_of_raw":"October 2026","source_type":"research_paper","comparison":null,"type":"estimate","figures":{"value":30,"unit":"%","approximate":false,"converted":null},"expiry_hint":null,"insight":null},{"id":"c15","statement":"Mercor suggested that the standard for measuring AI performance should shift from 'how well it performs tasks humans used to do' to 'whether it can accomplish tasks that were previously impossible or prohibitively expensive'","as_of":"2026-10","as_of_explicit":false,"as_of_raw":"October 2026","source_type":"research_paper","comparison":null,"type":"opinion","figures":null,"expiry_hint":null,"insight":{"analyst_name":null,"analyst_affiliation":"머코어","confidence":"high","forecast_horizon":"long_term","sentiment":"neutral","reasoning":"Since AI surpasses humans in standardized tasks, productivity measurement standards should shift from performance relative to humans to the feasibility of previously impossible tasks"}},{"id":"c16","statement":"It also forecast that even if model development stopped now, significant changes and productivity gains would still emerge in the accounting field","as_of":"2026-10","as_of_explicit":false,"as_of_raw":"October 2026","source_type":"research_paper","comparison":null,"type":"estimate","figures":null,"expiry_hint":"2027-12","insight":{"analyst_name":null,"analyst_affiliation":"머코어","confidence":"medium","forecast_horizon":"short_term","sentiment":"bullish","reasoning":"As standardized tasks shift to AI, productivity in the accounting field is expected to improve"}}],"ai_emotional_context":{"valence":0.2,"arousal":0.4,"primary_emotions":[{"emotion":"surprised","intensity":0.7},{"emotion":"concerned","intensity":0.5}],"secondary_emotions":[{"emotion":"skeptical","intensity":0.3}],"emotional_triggers":[{"claim_id":"c7","emotion":"surprised","reason":"significant_positive_metric"},{"claim_id":"c9","emotion":"concerned","reason":"significant_negative_metric"}]},"image":{"url":"https://onjblseywainslkhvkav.supabase.co/storage/v1/object/public/article-images/6d58c6d1-7abb-46b8-aa92-b60985f2cf28.png","alt":"정장 입은 인간형 로봇이 사무실 책상에서 서류와 계산기를 들고 데이터 화면을 보며 회계 업무를 처리하는 모습","caption":"(AI 이미지 제작)","source":"first_img","alt_status":"auto"}},"provenance":{"source_chain":["primary_reporting"],"original_source_url":null,"related_articles":[]},"temporal":{"freshness":"recent","next_update_expected":null},"access":{"license":"neupai_standard","attribution_required":true,"structured_data":"free","full_text_available":false,"full_text_access":null}}