Two numbers from this week tell opposite stories about the same technology. The first comes from the Center for AI Safety, whose Remote Labor Index measures how often AI agents can finish real, paid freelance projects, things like designing a 3D engagement ring, cutting a video ad, or mapping a floor plan, at a quality a client would actually accept. Anthropic's Fable 5 hit 16.1 percent, a record, and roughly double the next best model. As recently as eight months ago the whole field topped out at 2.5 percent. By this measure, economically useful AI is improving at a startling clip.
The second number comes from Torsten Slok, chief economist at Apollo, by way of Fortune. His charts show that outside a handful of tech and software firms, which can fold AI into their own products overnight, the profitability boost from AI simply is not appearing. Health care, banking, insurance, manufacturing, logistics, law, the public sector: the list of industries where the promised productivity gains remain invisible is long. If the technology is advancing as fast as the benchmark says, why can most of the economy not feel it?
The benchmark itself hints at the answer. Even Fable's record 16 percent is a long way from replacing a freelancer, and the researchers were careful about what the score does not mean. When they tried to hand the job of judging the AI's work to another AI, the model failed, because evaluating a deliverable means opening files in professional software and operating it competently, exactly the hands-on computer skills today's agents are weakest at. The same limitation that caps the score is the one that makes real-world deployment slow: a capable model still needs a human, and usually a whole scaffold of other tools, to turn its output into something a business can actually use.
This is the quiet tension running under the AI economy. Capability, measured in benchmarks, is racing ahead. Realised value, measured in profit margins outside the tech bubble, is crawling. Slok's worry is that the gap could become self-reinforcing: if non-tech companies keep failing to see a return, they may trim their AI budgets, and since token prices are already sliding toward zero for many tasks, there might not be enough revenue to sustain the enormous compute build-out the industry is betting on. Impressive models do not pay for data centres. Paying customers do.
There is an optimistic reading, and it has history on its side. Electricity took decades to reshape industry because factories had to be redesigned around it, not merely wired up to it. AI may be at the same awkward stage, powerful in the lab but bottlenecked by the unglamorous work of rebuilding workflows to use it. The freelance benchmark has quadrupled in eight months, and the computer-use skills that hold it back are precisely what the labs are now pouring money into. The miracle may simply be arriving on a slower clock than the demos suggest. Or the gap may be telling us something the benchmarks cannot: that being able to do the work and being trusted to deliver it are still very different things.