Claude Fable 5.2 completes 41.3% of assignments in the Remote Labor Index (RLI) with a result a paying client might accept. In the previous measurement, its predecessor Fable 5.1 had reached 17.9%; the best agent when the index launched, approximately a year earlier, scored 2.5%. The gap is striking, although perhaps the more useful figure is another one: even the model at the top of the table does not finish most of the evaluated jobs to a professional standard.
The RLI seeks to answer a question that knowledge tests or isolated programming tasks cannot resolve on their own: can an agent complete a project from start to finish under conditions similar to those of a paid assignment? The test brings together real jobs from different sectors and evaluates the delivered result, not just a brief response or an isolated skill. Fable 5.2's leap, therefore, does not amount to proof that AI can now generally replace independent professionals. It does show that a much larger share of certain assignments now reaches an acceptable threshold.
From 2.5% to more than 40%
The Remote Labor Index was created by Scale Labs and the Center for AI Safety. Its evaluation set contains 240 projects completed by 358 verified professionals on Upwork, with a combined value of approximately $144,000. The jobs include video editing, CAD design, game development, and data analysis. The average project costs $632.60 and takes a freelancer just under 29 hours.
This represents a change in scale from the starting point. The paper introducing the index on arXiv initially described results close to the floor: the highest-scoring agent automated only 2.5% of tasks. In the current measurement, Fable 5.2 reaches 41.3%. In July, Fable 5 was around 16%, and Fable 5.1 now sits just below 18%. The comparison sums up rapid progress, but it should not be read as if each percentage measured entire jobs that have disappeared: the indicator counts specific projects completed to a level a client would accept.
The difference also shows up in the current table. Anthropic puts Opus 5.5 at 21.3% and GPT-6 Astra at 20.8%; Fable 5.1 scores 17.9%. Fable 5.2 thus comes in at around twice the results of the next models in that ranking. Gemini Argon 4, according to The Decoder, scores 17.08%.
What counts as completed work
The agents receive a virtual workspace with professional software such as Blender, GIMP, and Audacity, as well as specialized tools for generating images, audio, and video. They then submit files that must be openable and usable. The evaluation does not merely check whether the result looks convincing in a screenshot: human experts review it in the relevant software and compare it directly with the work done by the original professional.
This process is a strength of the index and, at the same time, a bottleneck. According to The Decoder, attempts to automate the review with another AI model have failed: automated evaluators can be fooled by submissions that look polished but are technically defective, and award overly generous scores. Human inspection can detect these failures, but it makes evaluating many assignments labor-intensive. To interpret the 41.3%, the figure matters as much as the standard applied to decide what counts as success.
The variety of tasks also changes how to read the result. This is not a test of a single skill, but of complete assignments involving different tools and deliverables. Fable 5.2 improves especially in audio and music editing, architectural and interior design, and 3D modeling, according to Scale Labs. Nearly 50 of its completed projects were tasks that no agent in the evaluation had solved before. This points to advances in practical capabilities, although it does not prove that all sectors or types of work are progressing at the same pace.
The average conceals very different failures
The aggregate score does not mean that the system produces nearly finished results in the remaining 58.7% of cases. Failed attempts include incomplete work, broken files, contradictory designs, or amateur-level submissions. There are also areas where the model still struggles: Scale Labs highlights industrial product design and animation.
For a company considering bringing these agents on board, this mix matters more than their position in the table. A job that turns out well can save work; one that produces an unusable file may require review, corrections, and another run. The index records whether the work meets the professional standard, but the figure alone does not explain how much human effort would be needed to salvage each failed attempt. Nor does it say which specific tasks in an organization would match those the agent handles best.
The gap between an acceptable result and one that is merely attractive has direct implications for internal evaluation. A visual demonstration can conceal errors in the source file or in how it functions within the program. That is precisely why the RLI uses reviewers who open submissions in the relevant professional tools. For teams exploring task automation, the test suggests that measuring final artifacts against concrete acceptance criteria will be more informative than counting interactions or judging responses by their appearance.
An improvement with costs and conditions
Anthropic presents Fable 5.2 as a model for ambitious, open-ended technical work, and recommends starting with Opus 5.5 for many common programming and knowledge tasks. It also proposes that Fable can act as an adviser to smaller models or coordinate agents handling subtasks. This positioning matters: the RLI result does not force a choice between total automation and no use at all. It can help identify projects where the model can handle a substantial part of the work, with supervision and verification of the deliverable.
Cost also cannot be inferred directly from the success rate. Anthropic says Fable 5.2 costs less per million tokens than Fable 5.1 and produces results up to 30% faster. But Xataka, summarizing an external analysis by SemiAnalysis, notes that on GDPx the cost per task for Fable 5.2 was 31% higher than for Fable 5.1, because the model generated more tokens. That calculation does not concern the Remote Labor Index and does not allow us to conclude how much it costs to complete its projects; it does, however, warn that a lower unit price does not guarantee that each task will cost less.
Ultimately, 41.3% is a measure of demonstrated capability under a specific protocol, not an automation rate that can be applied as-is to any portfolio of work. The index features real assignments, professional tools, and human review, elements that make its results more like a work deliverable than those of many conventional benchmarks. But the remaining percentage is still the majority, and failures can range from an incomplete submission to a file that does not work. The next practical question for each company is not whether Fable 5.2 tops the table, but which assignments in its own workflow resemble those the model already handles successfully.
Sources: The Decoder, Xataka, Anthropic, arXiv
