AI can now do one in six freelance jobs

If you thought AI was about to replace every digital freelancer overnight, take a breath. But if you thought it had hit a wall, look again.

The Center for AI Safety (CAIS) and Scale Labs just updated their Remote Labor Index (RLI), and the numbers paint a fascinating picture of where autonomous agents actually stand. Instead of multiple-choice tests, RLI grades agents on 240 real-world freelance gigs—from 3D jewelry design and animated ads to architectural floor plans and audio engineering. Every file delivered by an agent is judged by human experts against work done by paid human pros.

The headline takeaway: agent capability in digital labor has surged nearly sixfold in less than a year, but the absolute ceiling remains stubbornly grounded.

Fable 5 pulls ahead of the pack

When RLI first debuted back in October 2025, GPT-5.2 led the board with an automation rate of just 2.5%. A few iterations later, Opus 4.6 managed 4.17%.

Now, Anthropic's Fable 5 has blown past previous benchmarks with a 15.8% automation rate (recorded at 16.1% in initial briefings), roughly double the 8.3% clocked by Opus 4.8. OpenAI's GPT-5.5 landed in third place at 6.3%.

Crucially, these agents weren't working in isolated sandboxes. CAIS and Scale hooked Anthropic models up to Claude Code and OpenAI models to Codex CLI, pairing them with full Linux virtual desktops packed with over 30 professional desktop applications, including Blender, FreeCAD, GIMP, and LibreOffice. The models took screenshots, clicked, typed, and exported real files just like a remote contractor would.

The automated judge gets caught bluffing

Scoring hundreds of complex freelance files with human experts is brutally expensive, so researchers attempted to build an LLM judge to grade the submissions. They tuned it on older models until its evaluations mirrored human scores.

Then things went sideways. When turned loose on the newer frontier models, the automated judge inflated GPT-5.5's scores by roughly 3x and Opus 4.8 by 2.5x. While it correctly ranked which model was better, it wildly exaggerated their absolute competence.

The reason is telling: judging creative work requires operating the software. A human client opens a 3D CAD file in Blender, navigates the viewport, and inspects the mesh geometry for structural flaws. The AI judge, unable to navigate desktop software reliably, looked at surface-level renders and assumed the work was fine. It inherited the exact same blind spots as the workers it was trying to evaluate.

Why it matters

We think this benchmark is one of the most honest assessments of the agent economy to date.

For freelancers and creative pros, the immediate takeaway is clear: your job isn't vanishing, but your workflow is shifting under your feet. Fable 5 beating human work nearly 16% of the time means roughly five out of six projects still fail client-level inspection when left entirely to the machine. You shouldn't fear being replaced by a prompt; you should be aggressively using these agents to handle first drafts and asset generation while keeping your hand firmly on the steering wheel.

For the AI industry, tripling the frontier automation rate in eight months proves that combining reasoning models with computer-use scaffolding is working. The progress curve is steep.

Most importantly, CAIS exposed the critical trap of using AI to grade AI. As models tackle complex, real-world work, automated evaluations decouple from reality. True quality control still requires human taste and professional rigor, and that bottleneck isn't going away anytime soon.