AI Agents Can Now Complete 16% of Freelance Work at Professional Quality

The fraction of freelance work that AI agents can complete at professional quality jumped from 2.5 percent to 16.1 percent in eight months — a sixfold acceleration that makes the displacement curve difficult to dismiss as a theoretical concern.

The Number

The Remote Labor Index (RLI) tracks 240 real freelance projects valued at $144,000 total, sourced from 358 verified professionals across 3D/CAD, graphic design, video production, audio, data analysis, and web development. Human evaluators from the Center for AI Safety assess whether each completed project meets the standard a paying client would apply. In the latest round, Fable 5 cleared that bar on 16.1 percent of assignments — the highest completion rate any model has recorded on this benchmark.

How the Math Works

Performance spread widely across models. Fable 5 led at 16.1 percent; Opus 4.8 followed at 8.3 percent; GPT-5.5 reached 6.3 percent; Gemini 3 Pro checked in at 1.25 percent. The gap between top and bottom is roughly 13×, which means model selection for production agentic pipelines is a performance variable — not a preference.

Fable 5’s number carries a documented caveat. Only 218 of 240 projects were fully evaluated before U.S. export restrictions temporarily removed the model from service. Even assuming failure on every remaining project, the floor would have been 14.6 percent — still ahead of all other models by a meaningful margin.

The benchmark’s most consequential finding is methodological: AI judges systematically overrated newer models by 2 to 3 times compared to human evaluators. The divergence was sharpest on tasks requiring correct operation of professional software — precisely where agents still struggle most. Benchmark scores produced using AI-as-judge evaluation should not be compared to RLI scores without a significant discount.

What It Signals

Sixfold growth in eight months makes the trajectory more interesting than the current percentage. At sustained improvement rates, capability thresholds for professional task categories shift from theoretical to operationally relevant within standard planning horizons. For context, agentic AI is already handling significant portions of pull request review work — a specific workflow domain where automation scope moved faster than broad benchmarks suggested. The RLI’s freelance domain tracking suggests a similar dynamic may be underway across creative and technical work categories.

The task-level view within the benchmark is also worth monitoring. The RLI spans domains with very different automation readiness profiles: data analysis and certain web development tasks have higher baseline automation rates than video production or complex graphic design. A 16 percent aggregate masks significant variance that matters for planning decisions at the discipline level.

Source: The Decoder

Content created with AI assistance and reviewed for accuracy.