OpenAI and Upwork Study: AI Projects Post 52% Bigger Budgets

OpenAI and Upwork published a joint study on November 10, 2026, and its sharpest finding is about project size. On Upwork, AI-related fixed-price projects post budgets about 52% higher than comparable non-AI work. Hired hourly AI contracts log about 41% more hours in their first 90 days. The posted hourly rate is only about 1.5% higher. So the extra money buys more hours and wider scope, and those bigger projects often combine tasks that clients rarely requested together before generative AI. The report is The Rebundling of Freelance Work in the AI Age by Caroline Chin, Alex Martin Richmond and Carl Shan of OpenAI and Shagun Kala, Nadav Kunievsky and Teng Liu of Upwork. OpenAI's summary is on its announcement page.

What did the OpenAI and Upwork study find?

Every figure below is quoted from OpenAI's announcement, the full report or the key takeaways PDF. "Adjusted" means the authors compared AI and non-AI posts in the same work subcategory and posting month.

I'd read the hourly rate line first. An adjusted gap of 1.5% means an hour of AI work is priced about the same as an hour of other work in the same subcategory. The bigger budgets come from more hours and wider scope. The authors put it this way: AI-related work "is associated with greater demand for labor and larger projects, rather than primarily with a higher price for each hour of worker time" (p. 9).

How is AI work bundled with integration?

The report's title is about rebundling: clients putting tasks into one posting that used to go to different specialists. Integration is one of the clearest cases, and the data says a few specific things about it.

Inside AI-related work, AI API integration gained share. Between 2023 and 2025, the tasks that gained share within AI-related postings were video production, social-media campaign management, marketing automation and AI API integration (p. 7). The authors describe demand shifting "toward integration (e.g., integrating AI APIs into applications), multimodal production... and applied business workflows" (pp. 2 to 3).

Integration in general did not grow. The same exhibit shows "data and systems integration" losing share within AI-related work, along with AI-generated written content, training programs and cloud infrastructure (p. 7). So the growth is in wiring AI into a product, not in integration work as a whole.

Chatbot work now tends to arrive with the build around it. Building chatbots and machine-learning models is still the largest cluster in AI-related work. Its share fell from 21.8% in 2023 to 17.8% in 2025 (p. 7). The report's examples show where that work went. Some postings "combine chatbot or machine-learning development with computer vision and AI API integration" (p. 11), and it cites "chatbot development combined with application integration" as a case of clients buying capabilities together "to obtain a finished outcome" (p. 14).

The new pairs are mostly AI tasks joined to traditional ones. In Appendix A, the authors sort the uncommon task pairs in AI jobs by type. Mixed pairs, with one AI task and one traditional task, are the largest group at about four per 100 pair occurrences, against roughly one to two for each other type (p. 16). And 98.6% of those uncommon pairs join two tasks that each already appeared in at least 200 pre-AI postings. The tasks themselves are familiar. Clients just hadn't put them in the same job before.

Integration bundles are new, though less so than creative ones. By type of AI work, generative AI and creative production shows the strongest effect: its task pairings had about 60% lower historical prevalence than comparable non-AI jobs. AI integration and automation is also statistically significant, but the gap is smaller (p. 13).

The report also limits its own claim. The mixed-pair result "should not be interpreted as showing that AI has broadly caused traditional tasks to be recombined with one another" (p. 16).

How was the study done?

• ChatGPT sample: 36,896 personal or work-related conversations from individual consumer ChatGPT accounts, sampled monthly from January 2025 to June 2026 (about 5,556 targeted per month). A classifier assigned each conversation to one of Upwork's work subcategories. Schoolwork and unclear conversations were excluded, and the authors state that no researcher read user messages (pp. 17 to 18).
• Upwork sample: Upwork Marketplace job posts. Category shares compare H1 2025 with H1 2026, and the AI intensity analysis uses H1 2023 as the baseline. Task-mix data runs through 2025. Budget, rate and hours comparisons cover posts through April 2026, and the top 1% of posted prices is trimmed (p. 18).
• What counts as AI-related: Upwork's internal GPT-powered classifier labels a post AI-related if it requests AI skills, requires AI tools or involves improving AI-generated output (pp. 7 and 17).
• Tasks: More than 3,000 activity descriptions grouped into more than 400 standardized tasks. The novelty regressions use 104,030 posts from a one-in-20 sample (pp. 13 and 20).

The caveats are the authors' own. The ChatGPT and Upwork datasets "are not directly matched, and neither represents the labor market as a whole" (p. 2). The price and hours comparisons "are descriptive and do not identify a causal effect" (p. 19). The 63-category chart "does not measure aggregate Marketplace growth or identify a causal effect of ChatGPT use on Upwork demand" (p. 18). So the study shows patterns inside one marketplace, and it doesn't claim AI caused the budget gap.

What does it mean for teams hiring AI-native engineers?

My read: scope the AI feature and the integration around it as one job, because that's how Upwork buyers increasingly post it and where the extra hours show up. Whoever you hire then has to own the model call, the app it lives in and the data it touches.

We place AI developers from our own bench in 2 to 4 weeks, from $3K a month, with a 90-day replacement guarantee. When the problem is still fuzzy and someone has to sit inside your team to map it, a forward deployed engineer embeds with you, builds evals on your real data and ships to production.

FAQ

Who published the OpenAI and Upwork freelance study?

OpenAI Economic Research and Upwork published it in November 2026 as part of OpenAI's Work at the Frontier series. The authors are Caroline Chin, Alex Martin Richmond and Carl Shan of OpenAI and Shagun Kala, Nadav Kunievsky and Teng Liu of Upwork. The full report is titled The Rebundling of Freelance Work in the AI Age.

Do AI freelancers on Upwork earn more per hour?

Barely. After controlling for work subcategory and posting month, posted hourly rates for AI-related jobs are about 1.5% higher than for non-AI jobs. The raw gap is about 9% (p. 9). The bigger difference is in hours. Hired hourly AI contracts log about 41% more hours in the first 90 days.

How much bigger are AI project budgets?

Posted fixed-price budgets are about 70% higher for AI-related jobs before adjustment and about 52% higher after controlling for subcategory and posting month (p. 9). These are posted budgets, not final payments, and the authors say the comparison is descriptive, not causal.

Is AI replacing freelance writers on Upwork?

The study points that way for content writing, but stops short of proving it. Content writing's share of Upwork postings fell by roughly 31% from H1 2025 to H1 2026, and the key takeaways say firms "may be using AI directly for this task." Across all 63 subcategories, the authors write that there is "not a simple substitution pattern" (p. 1).

What counts as an AI-related job in the study?

A job post that requests AI skills, requires AI tools or involves improving AI-generated output. Upwork's internal GPT-powered classifier makes the call using the title, description, skills and category of each post (p. 17).

What are the study's main limitations?

The ChatGPT and Upwork data are not directly matched, and neither represents the whole labor market. The price and hours comparisons are descriptive. Price data runs through April 2026, with the top 1% of posted prices trimmed. The ChatGPT sample covers individual consumer accounts only.