OpenAI and Upwork Find Bigger AI Projects, but Little Hourly Rate Premium

AI-generated editorial illustration.

On November 10, OpenAI and Upwork published three numbers that sound like a freelancer’s raise until their denominators are read. After comparing jobs in the same work subcategory and posting month, AI-related fixed-price projects on Upwork had 52.4% higher posted budgets. The adjusted difference in posted hourly rates was 1.5%. Among hired hourly contracts with positive recorded work, the AI-related jobs had 40.6% more hours in their first 90 days. The joint research report describes three different things: a proposed project budget, a proposed rate for a worker’s time, and hours recorded after a hire. None is a universal measure of what an AI freelancer takes home.

That separation changes the story. A client may be purchasing a more complete deliverable rather than paying much more for each hour. A freelancer may receive a longer engagement, but also take responsibility for integration, checking, revisions and client coordination. A fixed-price budget can rise even if the worker’s effective hourly earnings fall. The report does not calculate that effective rate for a comparable group of completed fixed-price contracts.

The researchers also found that AI-related job posts combined tasks in less familiar ways. A project that asks for AI-generated video, editing and motion graphics is buying a finished video. The draft produced by an AI tool is only one input. The buyer still needs someone to make the pieces work together and accept responsibility for the result. Yet broader briefs can be a bargain for the buyer and a burden for the worker, depending on how many revisions the agreement allows and who pays when the model’s output needs repair.

OpenAI and Upwork have a commercial interest in this market. Their study offers unusually detailed marketplace measures, but its findings are descriptive. It cannot establish that AI caused the wider briefs, that freelancers earned more, or that total employment increased. Independent research has found losses in some exposed freelance categories. The useful question for a worker is narrower than whether AI work is growing: after the extra production and review, how much paid work survives in a month?

Three prices in one job market

An Upwork client can post a fixed budget for a defined project or an hourly-rate range. The first is a ceiling or starting point for negotiating a deliverable. The second proposes a price for time. Neither is necessarily the final payment. For hourly contracts that reach a hire, the platform can observe recorded hours. The new report compares those three measures separately. OpenAI’s public account of the study emphasizes bigger projects and more hours; the technical report supplies the adjustments and the limits.

The unadjusted fixed-price budget gap was 70.3%. It narrowed to 52.4% after the authors compared posts within the same work subcategory and month. The unadjusted posted hourly-rate difference was 9.2%; after the same adjustment it was 1.5%. These comparisons do not hold project complexity, worker reputation, geography, client size, negotiated price or successful completion constant. The adjusted gaps are associations between AI-related classification and listed terms. They are not a causal estimate of an AI wage premium.

The report was released in November 2026, but its price and hours comparison uses jobs through April 2026. Its task-bundle analysis mostly uses postings through 2025. That lag is normal for research that classifies and compares large datasets; it also matters to an employer setting a contract this week. A client should use the paper to ask better questions about scope and rates, then check current offers in the specific category and market. The percentages are not a live quotation service.

The hours measure has an additional filter. It covers hourly postings that resulted in a hire and had positive recorded hours in the first 90 days. The adjusted AI-related difference was 40.6%. That is more work per selected contract, not a count of additional workers, a measure of all people who applied, or proof that every contracted hour was profitable. More hours can mean an expanded useful project. They can also reflect integration trouble or revisions. The study does not distinguish those mechanisms.

Published measure	Adjusted AI-related difference	What is observed
Fixed-price project budget	+52.4%	Budget stated in a posting, not final pay
Posted hourly-rate midpoint	+1.5%	Proposed rate range, not a negotiated worker wage
First 90 days of hourly work	+40.6%	Recorded hours on hired contracts with positive hours

The report trims the top 1% of the relevant posted-price distributions and compares posts in the same month and work subcategory. Those choices make the comparison more informative than a raw AI-versus-all-jobs average. They do not turn a platform posting into an earnings record. The hours analysis selects contracts that made it through hiring. A worker who bid on ten posts and won one appears differently from a worker who had ten paid contracts. The public analysis does not supply a worker-level estimate of that difference. It also excludes hourly contracts with zero recorded hours from the hours comparison. A signed agreement that never started may matter greatly to the worker, yet it is outside that estimate.

Upwork has published a much bigger pay-sounding number elsewhere. In its 2025 annual shareholder report, chief executive Hayden Brown said professionals working on AI projects earned 36% more per hour. The same letter said AI-related work exceeded $300 million annualized in the fourth quarter and was growing more than 50% year over year. These are company-reported commercial measures. The shareholder letter does not give the worker mix, adjustment method or time window needed to reconcile its 36% earnings figure with the new study’s 1.5% adjusted difference in posted hourly rates.

There need not be an error in either figure. The shareholder metric concerns professionals’ realized hourly earnings on AI projects, apparently across the platform’s worker mix. The joint paper concerns the midpoint of a rate range before a worker is hired and compares jobs within work categories and months. Experienced specialists might take more AI contracts; negotiations might change the posted rate; the mix of work may differ. Those are possible explanations, not findings established by the published materials. A buyer should not use 1.5% as a negotiating cap. A worker should not use 36% as a promised raise. The two measures ask different questions.

For a buyer, the larger posted budget may describe a larger bundle of work. For a freelancer, the decisive number is often the time and risk left after the contract is negotiated. A $1,000 project that takes 10 paid hours differs sharply from one that takes 25 hours and three unpaid rounds of revision. That arithmetic is illustrative, not a reported Upwork customer or a measured result from the study. The posted budget alone cannot tell those stories apart.

A bigger bundle, not merely a new label

The researchers used Upwork job text and its occupational taxonomy to construct more than 3,000 activity descriptions, grouped into more than 400 standardized task clusters. They classified AI-related posts with an internal language-model pipeline that considered the title, description, skills and work category. A job could be AI-related even when some of its component tasks were conventional production or review. The method is useful for studying combinations, but classification errors and changing language in postings remain possible.

Across AI-related postings, the effective number of equally represented task clusters rose from 14.6 in early 2023 to 21.2 in late 2025, about a 45% increase. That is a measure of how dispersed the mix of requested tasks became. It is not a statement that a typical job contained 21 tasks. Within individual postings through 2025, AI-related jobs averaged 3.35 distinct clusters, compared with 3.19 in other jobs. The within-job difference is modest, about 5%.

The more distinctive finding concerns which activities appear together. The report calls a task pair historically uncommon when fewer than 200 pre-AI postings contained it. Among current jobs with at least two tasks, the average share of such pairs was 8.4% in AI-related posts and 2.4% in other posts. Those are within-job shares, not shares of all jobs. The authors also compare posts with similar bundle sizes in the same work subcategory and month; AI-related projects still contain pairings with less precedent. The result is a shift in composition beyond simply adding more tasks to every brief.

The 200-post threshold is a research definition, not a claim that the pairing was invented by AI. A rare combination could have existed in another marketplace, inside an employer or in an unrecorded subcontract. The analysis is most useful for identifying a change in how clients describe purchased work on this platform. It cannot tell a training provider that every uncommon pair deserves a new certificate, or tell a worker to abandon a specialty because it appears less often in a sampled posting.

Consider a video assignment. The report describes AI-generated video combined with editing and motion graphics. A client can specify a generated first pass, ask for a finished cut, then require motion elements that fit the campaign. A freelancer who once sold only editing may need to know how to produce, select and correct generated clips. Another worker may specialize in the final polish. The platform posting cannot tell whether one person completes all three steps, a team shares them, or a buyer writes a broad brief and later narrows it. It does show that those activities are being requested together.

AI-generated editorial illustration. The branches explain possible project scope, not a measured client workflow or a chart of the study.

That distinction matters when hiring. A brief that combines new tools with conventional production can reduce handoffs if one experienced person owns the final output. It can also hide several specialists’ work behind one fixed budget. An employer or procurement team should ask which outputs must be delivered, what quality standard applies, who reviews the AI-generated part, whether source material can be used with the tool, and how many revisions are included. A worker should ask the same questions before treating a larger project amount as higher pay.

The task taxonomy is another boundary. A language model helps identify tasks, then a clustering method and manual review place them into a stable set of groups. Such a map can reveal broad changes, but it cannot inspect every client’s real implementation. A post may list a tool as a preference, not a requirement. Some employers use the term AI to attract candidates; others use it without changing the job. The published report does not provide an external audit of every classified posting. Its strongest evidence is the observable mix of posted descriptions and recorded hours, not a complete account of how each project was delivered.

There is an important difference between a larger bundle and a better occupation. A worker who can integrate a model into a business process may command a longer contract. A specialist whose former task becomes a line item in that contract may lose direct access to the buyer. A client could split the work among specialists, hire one generalist, or keep the easiest step in-house with an AI tool. The posting text records the requested bundle; it does not name the allocation that followed. The distribution of bargaining power sits behind the same task count.

Direct AI use and paid work diverge

The report compares two different datasets. One samples consumer ChatGPT conversations between January 2025 and June 2026. The other examines Upwork Marketplace job posts. ChatGPT users may ask for writing, advice or planning for themselves. Upwork clients post work they want another person to do. The populations are not matched. Their geographic mix, ability to pay and reasons for using each service differ.

Writing illustrates the contrast. In the classified ChatGPT sample, content writing represented 14.5% of conversations assigned to an Upwork-like activity, while professional and business writing accounted for 10.2%. In Upwork posts, the comparable shares were about 1.3% and 0.4%. This does not mean ChatGPT eliminated writing contracts: the denominators differ, and a conversation can be a small step within an ongoing job. It does show why counting prompts is a poor proxy for purchased freelance work. People often use a general-purpose assistant for a quick draft or advice that they might never have bought as a standalone service.

The researchers compared 63 matched work subcategories, measuring ChatGPT intensity against changes in Upwork category shares. 43 fell into the expected directional pattern: activities relatively common in ChatGPT lost Upwork share, or less common activities gained it. The other 20 did not. A share change is not an absolute count. If one category grows more slowly than the entire marketplace, its share can fall even while its posting volume rises. The study’s chart does not establish total marketplace demand or a causal effect of ChatGPT.

This makes a buyer’s make-or-buy decision less tidy than an automation forecast. A marketing manager may use ChatGPT to outline a campaign, then hire an editor to test claims and produce assets. A software team may ask an AI tool for a code sketch, then hire an integrator for permissions, deployment and maintenance. Those are illustrative decisions, not cases observed by Digidai. In both, the direct AI use and the freelance contract can coexist. Other clients may decide the draft is enough and never post the job. The report cannot count all of those unposted alternatives.

Upwork’s own earlier analysis of AI-remediation postings found that posts asking people to fix, refine or finish AI-generated output had risen 70% year over year on a trailing 12-month measure through the second quarter of 2026. The company sampled half of posts in twelve marketplace categories between January 2023 and June 2026. This is a platform demand signal, not a count of hires or proof of rising earnings. It is also a reason not to reduce every AI-assisted project to the first draft. Someone still has to decide whether the result works.

The worker who does not win the project

Independent evidence keeps the new report from becoming a simple story of expanding opportunity. Ozge Demirci, Jonas Hannane and Xinrong Zhu’s 2025 study in Management Science found that postings for automation-prone writing and coding jobs fell 21% relative to manual-intensive jobs within eight months after ChatGPT’s release. The authors also found a 17% relative decline in image-creation postings after image generators became available. Remaining exposed jobs were more complex and offered higher pay. That last finding can coexist with fewer entry opportunities. A higher budget among surviving jobs does not compensate a worker who receives no contract.

Xiang Hui and Oren Reshef, writing about their platform research for Brookings, reported about 2% fewer monthly contracts and 5% lower monthly earnings for freelancers in more exposed occupations after early generative-AI releases. Their setting and periods differ from the new OpenAI-Upwork comparison. The point is not to average incompatible percentages. It is to separate a posting’s proposed budget from a person’s realized portfolio of paid work.

The older studies are not a rebuttal of every new opportunity either. They examine the early shock after generative tools appeared, while the joint paper examines later combinations of AI and production work. Different clients could have stopped buying simple tasks and begun buying integrated projects. The same client could even do both. To show that the new projects offset lost work, though, researchers would need to follow workers across categories and count their net earnings, including those who left the platform. A larger surviving project is not a replacement contract for everyone who used to sell a smaller task.

Selection can produce all three findings at once. Suppose routine writing contracts disappear while the remaining ones demand research, editing and distribution. A researcher could observe fewer contracts for workers, larger budgets for posted projects, and more hours on the hired contracts that remain. Whether that happened at scale in 2026 requires linked evidence on applicants, hires, negotiated terms, time and earnings. The joint report does not provide a full panel of worker outcomes. The older studies do not tell us whether the latest AI-related project bundles will persist or how they are distributed across workers.

There is a geographic boundary too. Upwork is a global digital marketplace, and the ChatGPT comparison samples consumer accounts. U.S. Bureau of Labor Statistics data come from a household survey about work arrangements, including independent contractors on a person’s main or secondary job. A platform job post and a household’s main job are not interchangeable units. Upwork’s July 2026 survey estimate that 38% of skilled U.S. knowledge workers freelance has a narrower population and a different methodology from official independent-contractor measures. It should not be used as a U.S. labor-force rate.

A contract can put the risk in the wrong place

The new report gives a buyer a reason to examine scope before selecting a rate. AI may lower the cost of making a draft, but a deliverable can acquire new steps: checking rights to input material, verifying facts or code, integrating an output into existing systems, and testing it with people. Some of those steps are billable. Others can slip into the fixed price as invisible labor. A posted budget says nothing about which side bears the cost of a wrong result.

Consider an AI-enabled customer-service workflow. A buyer could ask a freelancer for a bot prototype, a knowledge-base connection, evaluation prompts, escalation rules and a handover document. That is a plausible brief, not a documented Upwork posting in this report. If the contract says only “build an AI assistant,” the buyer may assume support and testing are included while the freelancer has priced a prototype. The job becomes larger without a clear agreement about what finished means. The same problem appears in design, video and marketing production.

An hourly contract makes some extra work visible, though visibility is not a guarantee of payment. The report’s 40.6% adjusted hours gap applies only to hired contracts with recorded positive hours. It cannot say whether the worker spent additional unrecorded time finding the client, learning a tool or redoing rejected material. A fixed-price contract may hide all three in the worker’s effective hourly rate. To compare offers, a freelancer needs the negotiated fee, expected hours, number of revisions, platform fees and probability of payment. The joint study gives public readers only some of those pieces.

Upwork’s own fixed-price guidance describes milestones as separate deliverables with separate payment amounts. Its client payment guide says a client has 14 days to review submitted work, approve it or request changes; absent action, the platform treats the submission as approved and releases payment. Those operating rules matter more to a project than the headline percentage. A worker can insist that model integration, review and final delivery are named milestones before agreeing to the fee. A buyer can fund the next stage after the first output passes an agreed test.

That arrangement does not remove disputes or guarantee a return. It makes the point of disagreement visible. If the AI-generated video passes the first storyboard review but fails a later rights or brand check, the parties can see which deliverable was accepted, which work remains, and whether the new request changes the scope. Without that record, the larger fixed price can absorb an unlimited series of small corrections. The report does not measure how many projects use milestones, how often they are revised or how much unbilled rework occurs. It provides a reason to collect those facts.

The client has a measurement problem of its own. A larger budget can buy a more integrated output, but it can also buy a longer experiment that never reaches production. Compare the accepted deliverable with the original scope. Track revision rounds, total paid hours, the handoffs to other workers, and whether the project is used after delivery. Paying for more hours is sensible if the result is worth it. The posting itself does not establish that value.

The unit of success should be specified before a buyer posts the project. For a video, it might be an approved cut in the requested format with rights cleared for the intended channels. For software, it might be a working integration that passes a small set of tests and can be maintained by the client’s team. The agreement should say who supplies inputs, who checks output, and what happens when the source material changes. These conditions can make a high budget reasonable. Their absence can turn the same budget into a dispute about unfinished work.

Follow the worker through the funnel

The November report is strongest where it sticks to the units it can observe. AI-related Upwork jobs show a broader mix of tasks and historically less familiar combinations. In comparable categories and months, they state larger fixed-price budgets. Hired hourly contracts record more work over 90 days, while adjusted posted hourly rates are close. The findings make a larger-project hypothesis plausible. They do not show that AI raised a typical freelancer’s wage or that the total pool of paid work expanded.

A future comparison should keep five denominators separate. Count posted briefs to see what buyers requested. Count awarded contracts to see which requests became paid work. Record the negotiated fee or rate to see what the parties agreed. Record paid and unpaid hours to estimate the worker’s return. Finally, follow the same worker over time to learn whether added AI work replaced lost assignments or increased total earnings. Each denominator can move in a different direction. A marketplace can advertise more ambitious projects while a particular worker faces fewer chances to win one.

The next useful evidence would follow the same worker across the whole funnel: postings seen, bids sent, contracts won, final prices, billed and unbilled time, revisions, disputes and earnings. Break that record down by entry-level and established workers, by location, and by the activities most exposed to direct AI use. Pair it with client outcomes. That would show whether broader bundles create durable work or merely move more responsibilities into fewer contracts.

For now, the decision sits in an ordinary project brief. A buyer can write down what the finished work must do and pay for the steps required to get there. A freelancer can price those steps and reject a budget that shifts all the uncertainty onto them. OpenAI and Upwork’s 52.4% gap is a useful prompt to inspect the brief. The invoice and the worker’s hours will determine what the larger number meant.