Eight months ago, the toughest test yet devised for working AI delivered a humbling verdict: the best agents on the market could complete just 2.5% of real freelance projects to a paying client’s standard. This week, the Remote Labor Index — the benchmark behind that number, built by the Center for AI Safety and Scale AI — published updated results, and the picture has changed dramatically. Anthropic’s new Fable 5 model completed 16.1% of the projects at a level a client would accept. That is roughly one in six real jobs, and the fastest climb the benchmark has recorded.
What the Remote Labor Index actually measures
Most AI benchmarks test puzzles, exams or code snippets. The Remote Labor Index is different: it is built from 240 complete, real projects sourced from freelance platforms such as Upwork, spanning 23 professional categories — 3D jewellery design, animated ads, architectural floor plans, game development, data work and more. Together the projects represent over 6,000 hours of human labour worth more than $140,000, with a median project taking a professional about 11.5 hours and paying around $200.
The grading standard is the one that matters commercially: would the client accept this deliverable and pay for it? Not “is the answer plausible” — but “is the work done”. Full methodology and current standings are published on the Scale AI leaderboard.
From 2.5% to 16.1% in eight months
When the index launched in October 2025, the leading model automated 2.5% of projects. The July 2026 update puts Fable 5 at 16.1%, with Claude Opus 4.8 second at 8.3% and OpenAI’s GPT-5.5 third at 6.3%, according to the Center for AI Safety’s analysis. A sixfold improvement in under a year is the headline — but the shape of the results matters just as much.
The models succeed most on generative work: producing images, audio, copy or code from a clear brief. They still fail most often when a project demands precise multi-step specifications, heavy tool use, or careful editing of existing material — exactly the messy middle where most real client work lives.
What this means for a small business
The practical takeaway is not “AI will replace freelancers next quarter”. It is that a meaningful — and rapidly growing — slice of routine, well-specified digital work can now be delegated to AI, provided someone competent writes the brief and checks the output.
Three implications stand out for owners and consultants:
1. The brief is the bottleneck. The projects AI completes are the ones with clear specifications. Tightening how work is scoped and documented pays off twice: better human outcomes and more AI-delegable tasks. This is a core discipline of good AI project governance.
2. Treat AI like a junior freelancer, not an oracle. A one-in-six acceptance rate means quality control is non-negotiable. AI output should pass through the same review step a new contractor’s work would, with results tracked the way a business would measure any AI investment’s ROI.
3. Revisit the boundary every quarter. The 2.5%-to-16.1% jump shows how quickly the frontier moves, especially as always-on AI agents get better at long, multi-step work. A task that failed in January may be routine by October. A short, regularly re-tested list of “not yet” tasks keeps the boundary visible.
Limitations and what to watch
The number worth remembering is the inverse: even the best model still fails five projects out of six by a paying client’s standard. The Center for AI Safety’s own commentary notes that current systems still fall short of professional quality on most work, and that even some deliverables from the top-scoring model would not be accepted as finished output. The index also measures fully autonomous completion — a single agent handed a brief — which understates what human-plus-AI workflows achieve and says little about work requiring client conversations, revisions or accountability. Finally, one benchmark, however realistic, samples only a slice of the freelance economy; categories outside its 23 project types may automate faster or slower.
The bottom line
The Remote Labor Index is the closest thing the industry has to an honest scoreboard for AI doing real, paid work — and the score just moved from “barely” to “sometimes”. For small businesses, the winning posture is neither panic nor dismissal: build the briefing, review and measurement habits now, so that each time the number climbs, the savings are captured early.
