AI Agents: From 2.5% to 16% of Freelance Work in Eight Months

Eight months ago, AI agents could complete 2.5% of real freelance projects at a quality level clients would actually pay for. Today that number is 16.1%—roughly a 6× increase, driven almost entirely by frontier model improvements.

The Number

The Remote Labor Index (RLI), published by the Center for AI Safety, benchmarks how frequently AI agents complete commercially viable freelance work at professional quality—meaning clients would accept the deliverable. It’s not a theoretical capability score. It’s a pass/fail on whether real work gets done at a standard humans would pay for.

The dataset behind the benchmark: 240 projects worth $144,000 combined, sourced from 358 verified freelancers. Job categories span 3D and CAD work, architecture, graphic design, video and animation, audio, data analysis, and web applications.

Current model performance:

• Fable 5: 16.1%
• Opus 4.8: 8.3%
• GPT-5.5: 6.3%
• Gemini 3 Pro: 1.25%

The spread between models is enormous. Fable 5’s 16.1% is more than twice Opus 4.8’s rate. Whatever Anthropic did architecturally with Fable 5—and the model’s general-purpose reasoning scores suggest a step change from the prior generation—it’s reflecting in real task completion, not just benchmark performance.

How the Math Works

The rate of 16% matters less than the trajectory. Going from 2.5% to 16.1% in eight months isn’t linear improvement—it’s a compound dynamic. Each capability threshold crossed unlocks additional task categories.

There’s a critical caveat buried in the benchmark methodology: AI evaluators rated deliverables “almost three times too high” compared to human evaluators on GPT-5.5. The implication is significant. AI judges lack the operational context to open and assess files in the right professional software, or to catch subtle quality gaps that professionals recognize immediately. This means the 16% figure represents human-evaluated quality, not AI-evaluated quality—which would be inflated considerably higher.

What This Signals

The RLI is a lagging indicator by design. It measures completed freelance categories, not what’s emerging in adjacent areas. The jump from 2.5% to 16% should be understood as the floor, not the ceiling, of what’s happening in AI-driven labor displacement.

More practically: the gap between the top-performing model (Fable 5 at 16.1%) and everything below it suggests that benchmark ceilings are being hit by a specific class of model—not the field broadly. Teams selecting models for autonomous work pipelines should weight this benchmark heavily relative to general-purpose reasoning leaderboards, which don’t capture the multi-step, context-switching nature of real freelance projects.

Source: The Decoder

Content created with AI assistance and reviewed for accuracy.