Claude Fable 5.2 Matched a Professional on 41% of Freelance Projects: What the Number Measures

Anthropic released Claude Fable 5.2 on 9 November 2026. One row in the benchmark table of its announcement concerns paid work directly: a score of 41.3% on the Remote Labor Index, against 21.3% for Claude Opus 5.5 and 17.9% for Fable 5.1, the version it replaces. The system card published the same day puts the result in a sentence, saying the model “completed 41.3% of freelance projects at least as well as a human professional.”

A sentence like that is easy to shorten in a planning meeting to “AI can do 41% of the work” or “41% of jobs.” Neither version matches what was tested. This piece explains what was measured and what was left out, then sets out how an HR or finance leader can put the result to use.

What is the Remote Labor Index?

The Remote Labor Index is a benchmark, meaning a fixed set of test items that different AI systems attempt so their results can be compared. Researchers at the Center for AI Safety and Scale AI built it and described it in a paper dated 30 October 2025. The test items are 240 projects completed by human professionals, most of them collected from professionals with verified Upwork accounts who sold the researchers samples of past work.

Each project has a written brief, the input files the freelancer worked from, and the deliverable the freelancer handed over. The paper’s examples include an animated advertisement for a tree services company and architectural plans for a container home. Video, computer-aided design, graphic design, game development and audio are the largest categories, at 10% to 13% each, and the set covers 23 of the 64 work categories Upwork uses.

The paper reports a median cost of $200 and a median completion time of 11.5 hours per project, as stated by the professionals who did the work. A few projects are far larger, with costs above $10,000 and completion times above 100 hours. Taken together, the authors put the set at more than 6,000 hours of professional time and $143,991 in fees.

How was “matched a professional” judged?

People did the judging. The AI agent receives the brief and the input files and produces its own deliverable. According to the paper, three trained evaluators then each compare that deliverable with the professional’s and decide whether it “would be accepted by a reasonable client as the commissioned work.” The majority view decides the outcome for that project. Each evaluator was given up to 20 minutes for a project and took 11.4 minutes on average, and the evaluators agreed with one another 94.4% of the time.

The headline figure is what the authors call the automation rate: the percentage of projects where the AI deliverable was judged to complete the project at least as well as the professional’s. For Fable 5.2 the system card gives this as 99 of 240 projects. It also says Scale AI ran the test at high effort with a 200,000-token context window, settings the card describes as defaults. Scale AI is one of the benchmark’s two creators, and the same 41.3% appears on its public leaderboard in an entry dated 9 November 2026. That makes this one of the results in the system card that an outside party produced.

What does the number leave out?

The unit is a project. A project is a different thing from a job. A job is a continuing set of tasks, done with colleagues, for clients and managers who change their minds. The paper’s authors say they organized the benchmark by Upwork’s categories because the federal O*NET occupation list was a weaker fit for freelance projects. They also state the limit plainly: a system that scored 100% could still underperform people on the kinds of work the benchmark does not test (paper, section 5).

Adoption is measured separately. A benchmark shows what a system produced under test conditions, and usage data shows what people do with it at work. Anthropic’s own labor market study from March 2026 uses a measure it calls observed exposure, which combines estimates of what the models can do with real usage of Claude. On that measure computer programmers ranked highest, with 75% of their tasks covered, and 30% of workers were in occupations with no measured coverage, which means too little usage to register and says nothing about future potential. The study found “no systematic increase in unemployment for highly exposed workers since late 2022.”

The result describes the past. When the benchmark launched in October 2025, the best agent scored 2.5% (paper). A little over a year later the best score is 41.3%. That is a large change, and it describes two measurements in the past. Anthropic’s announcement does not mention jobs or the labor market, and the system card draws no projection from the result.

The decimal overstates the precision. With 240 projects, each one moves the score by about 0.4 points. The leaderboard publishes no margin of error and states: “There is no confidence interval for Automation Rate.”

What Anthropic says about the model, and where its own results differ

Anthropic’s announcement does not position Fable 5.2 as the default choice for office work. It recommends that users start with Opus 5.5 for most coding and knowledge tasks and reserves Fable for “the most ambitious, open-ended technical work.” The model documentation lists Fable 5.2 at $8 per million input tokens and $40 per million output tokens, twice the Opus 5.5 rate.

Other benchmarks in the same system card do not all point the same way. On GDPval-AA v2.1, a test of 220 professional tasks across 44 occupations run by the firm Artificial Analysis, Fable 5.2 received an Elo rating of 1831 and trailed Opus 5.5 at 1866. An Elo rating ranks models against each other from head-to-head comparisons of their output, and a higher number is better. On an internal finance evaluation of 123 tasks, Fable 5.2 did not finish six within the three-hour limit, while Opus 5.5 finished all of them. On the question of replacing its own researchers, Anthropic writes that Mythos 5.2, which the announcement describes as the same underlying model, “does not seem close to fully substituting for our research staff,” and labels that assessment uncertain.

All of these figures come from a vendor’s document about its own product. According to the system card, the Remote Labor Index and GDPval-AA runs were carried out by outside organizations, and the finance evaluation was built and graded inside Anthropic.

How should an HR or finance leader read a result like this?

Read it by task first. The benchmark tests one shape of work: a deliverable that one person can produce alone from a written brief and a folder of files. For each role, separate the work of that shape from the work that is different, such as meetings, negotiation, supervision and decisions that depend on context nobody wrote down. The useful question is how much of the week goes to brief-and-deliverable work, and the answer for a motion designer is unlikely to match the answer for an account manager.

Then read it by role and by budget line. The projects in the test were bought on a freelance marketplace, so the nearest match inside a company is outsourced project work: the design, video, drafting and analysis jobs that go to freelancers and agencies. That spend is a more direct comparison than payroll. For employees, the result is a reason to review which tasks in a role may change. It is weak evidence for removing the role, a point our guide Before You Cut Headcount for AI covers in detail. JobRoute’s own task scores draw on observed AI use from the Anthropic Economic Index together with O*NET task data, and model benchmarks are not an input (methodology).

Include review time in the cost. A 41.3% pass rate (system card) means the other 58.7% of deliverables were judged short of the professional standard, and nobody knew which were which until trained people had checked every one. In the benchmark each project was checked by three evaluators who averaged 11.4 minutes apiece (paper). In a company, record who does that checking and how long it takes, because that time belongs in the business case alongside the model’s usage fees.

Test on your own work before changing a plan. Pick ten or twelve past briefs where you still hold the files and the deliverable you accepted. Have an AI agent attempt each one, and ask reviewers who do not know which version is which to apply the benchmark’s standard of whether a reasonable client would accept it. The result will be a small sample, and it will describe your work and your quality bar. Put a date on it, because the score on this benchmark rose from 17.9% to 41.3% between one version of the model and the next (system card).

For the wider evidence on exposure and employment, see AI and Jobs in 2026: What the Data Actually Says. If a task-level view of your own roles would help, JobRoute builds a free AI exposure sample for 10 roles, and you can request it here.