A new AI finished 41% of real freelance jobs on a test | The Receipts - Dataspheres AI

A new AI finished 41% of real freelance jobs on a test THE RECEIPTS A new AI finished 41% of real freelance jobs on a test November 11, 2026 Anthropic rele...

A new AI finished 41% of real freelance jobs on a test THE RECEIPTS A new AI finished 41% of real freelance jobs on a test November 11, 2026 Anthropic released Claude Fable 5.2 on Monday Anthropic released Claude Fable 5.2 on Monday — An AI model is software that writes, codes and makes files when asked — In the company’s video, on screen, the model steers robot arms Robot arms steered by the model pick up a puzzle cube, in Anthropic's announcement video Anthropic, from its November 9, 2026 announcement video It finished 41.3% of the real freelance projects on a test It finished 41.3% of the real freelance projects on a test The projects are real paid jobs The projects are real paid jobs — 240 projects that freelancers were paid to finish — Each has the client’s brief, the files and the finished work — Dashboards, product animations, games, building plans, adverts A real paid project from the test: a dashboard built from a spreadsheet, with the work the freelancer delivered A real paid project from the test: 3D animations of a new earbuds design, with the work the freelancer delivered A real paid project from the test: a small web game, with the work the freelancer delivered One job from the test: plans for a container home One job from the test: plans for a container home — The client asked for plans and a 3D model of a container home — The AI gets the same brief and the same files — It has to hand back finished files, not advice — The freelancer’s delivered work is the bar One company made the model, two outside groups built the test One company made the model, two outside groups built the test — Anthropic made the model — The Center for AI Safety and Scale AI built the test — The company that made the model did not build the test The logo of Anthropic, the company that made the model The logo of the Center for AI Safety, a research group that built the test The logo of Scale AI, which built the test with it and publishes the leaderboard The test in numbers The test in numbers 240 real projects, from 358 freelancers on Upwork, a hiring website $632.60 what the average project was worth $143,991 what the projects were worth in total 28.9 hours average time a freelancer spent on one People grade every job by hand People grade every job by hand People grade every job by hand People grade every job by hand People grade every job by hand The top score has doubled about every 3.3 months The top score has doubled about every 3.3 months Oct 2025 The test launches. The best AI finishes 2.5% Nov 24, 2025 Opus 4.5 reaches 3.8% Apr 23, 2026 GPT-5.5 reaches 6.3% Jun 9, 2026 Fable 5 reaches 15.8% Sep 3, 2026 GPT-6 Astra reaches 20.8% Nov 9, 2026 Fable 5.2 reaches 41.3% The next best system finished 21.3% The next best system finished 21.3% — Opus 5.5 is from the same company, Anthropic — GPT-6 Astra is from OpenAI The company says it built an operating system in 8 days The company says it built an operating system in 8 days — The company says more than 1,000 AI helpers worked around the clock for 8 days — It starts up on a laptop with windows and a login screen A laptop running the prototype operating system Anthropic says its model wrote, in the company’s announcement video Anthropic, from its November 9, 2026 announcement video The company says it designed three gadgets that worked The company says it designed three gadgets that worked — A voice-assistant hub, a USB speakerphone and a handheld virtual pet — All three worked with no electrical fixes, the company says A circuit board for one of three devices Anthropic says its model designed, in the company's announcement video Anthropic, from its November 9, 2026 announcement video On a robot test the company ran, its score more than doubled On a robot test the company ran, its score more than doubled — In the company’s video, robot arms steered by the model pick up a cube and draw — 64.9% on a simulated robot test the company ran itself, up from 27.6% for Fable 5.1 In the company’s video, two robot arms steered by the model pick up a puzzle cube In the company’s video, a robot arm steered by the model draws a picture of a robot Most of the jobs were still not finished Most of the jobs were still not finished — 58.7% of the projects did not meet the standard — The test covers 23 of the 64 kinds of work on Upwork — It leaves out physical work and work done with a client — It leaves out projects that take a long time to judge It is still weak at animation and product design It is still weak at animation and product design — Still weak: industrial product design and animation, The Decoder reports — Biggest gains: audio and music editing, architecture, 3D models — Nearly 50 of its finished jobs were firsts for any AI on this test The price fell, and one task can still take most of an hour The price fell, and one task can still take most of an hour $40 per million tokens of output, down from $50 30% faster at most, the company says 44.4 min per task on an outside index $26