Fable 5.2 delivered acceptable results on 41.3% of freelance projects in RLI

You ask an AI to prepare a 3D scene for a product listing. An hour later, the agent reports that everything is ready. But does the project open? Can you change the material of the casing? Do the dimensions match the brief? The promise of automation often breaks down in precisely this gap between a polished response and usable work.

In Remote Labor Index, Fable 5.2 passed 41.3% of projects. Compared with Fable 5.1's 17.92%, that is an increase of 23.38 percentage points. For a developer or founder, this offers a concrete reason to experiment: find out which complete assignments can already be handed to an agent under a specialist's supervision.

What the 41.3% result means

On the Scale Labs leaderboard, Fable 5.2 leads Opus 5.5 at 21.25% and GPT 6 Astra at 20.83%. That puts it almost twice as high as its closest competitors. But the percentage itself needs careful interpretation: it is the share of projects that passed within a particular set of tasks and evaluation conditions.

Experts evaluate the agent's work by comparing it with a human professional's deliverable. The key question is whether it meets the brief well enough for a reasonable client to accept it. This comparison brings the evaluation closer to teams' everyday work, where demonstrating a single successful step is not enough.

Still, 41.3% should not be read as the probability that any given commission of yours will succeed. Your team may have different requirements, formats, and task distributions. And even on the benchmark itself, most projects still do not pass.

How to test this in your own workflow

Let us return to the scene for a product listing. This is a hypothetical example, not a published result from a specific Fable task. Suppose the commission includes an object model, lighting, several camera angles, and an editable source file. An attractive image is not enough for acceptance: the next designer still needs to work with that file.

A practical pilot should focus on the complete delivery package. Give the agent the same brief and source materials that a contractor would receive. Write down the criteria in advance: required dimensions, file contents, editability, and alignment with references. Then hand the result to a specialist who knows how to assess this kind of work.

Ask the reviewer to record the time spent on corrections separately. An agent may quickly produce a starting point but require extensive manual refinement. For a business, those hours are part of the cost of the result, alongside the expense of running the agent and reviewing its work.

It is also useful to retain failed attempts. If the system regularly loses textures or overlooks some requirements, that gives you a concrete engineering problem: add a check that all required files are present, clarify the inputs, or narrow its scope of use. A single average score is not enough to make that decision.

Where the leaderboard's conclusions end

The original Telegram post separately claims advances in audio, architecture, and 3D modeling, as well as nearly 50 tasks solved for the first time by any system. The public leaderboard page does not substantiate these details with a separate analysis of Fable 5.2. It is too early to use them as a basis for choosing a tool.

The comparison itself also has a limitation: RLI excludes projects that require direct communication with the client. In a real commission, aligning expectations, clarifying an ambiguous requirement, and responding to revisions can account for a substantial share of the work. The test result therefore does not describe the full client relationship lifecycle. The RLI methodology also states that the Automation Rate metric has no confidence interval.

For a product manager, a sensible next step is to select a repeatable type of assignment with clear acceptance criteria and run a small internal pilot. Adoption decisions should be based on the share of accepted deliverables, specialist time, and total cost. The rise to 41.3% makes that test more compelling; teams still need to measure the economic impact themselves.