2.6 "Best to median" sounds like a much more plausible figure. The x10 figure is usually quoted in such a way that it seems like it is "best to median".
Projects are sized based on a negotiation between us and the customer. It is then broken into parts, which are given their own kousuu value, via guesstimation by a senior engineer on the team. (You'll notice this process involves an awful lot of guesstimation. This tends to converge on reasonable numbers because guesstimation is happening among multiple actors who don't have an incentive to collude -- kinda-sorta like a market of players with imperfect knowledge can converge staggeringly quickly on a good guess of the intrinsic value of something.)
We then measure how many kousuu of work you complete for a particular project per day you report you are working on the project. That is currently our primary productivity measure, although we also track customer satisfaction, "claims" (a backchannel by which customers here get free work done to fix issues in the deliverable which are determined to be due to our mistakes), subjective quality assessments from the engineers, follow-on business from the same customer, and any value we can extract from the project for the company. (For example, in a 10 kousuu project if we manage to extract 2 kousuu worth of functionality that the customer paid for into our internal framework, then deploy that to seven customers and charge them each for it, we make out like bandits.)
Anyhow, we keep obsessive records of who does what and how long it took, so figuring out productivity is "just" a matter of dividing. (We only figured out the dividing trick recently. Hard to move a large organization from the way it typically does things, you know?)
1. My source for the productivity figure is Peopleware. They cite a set of "coding war games" that found a repeatedly measured 10x variation between the best and the worst programmers in their work environment. The productivity difference showed up under a variety of different measurements. There is no vagueness about the measurement techniques or doubt about the figure.
2. The 10x figure was from best to worst, from best to median was (IIRC) 2.6 times.
3. A significant portion of the figure they measured was correlated with environment, and the strongest predictor of performance was the performance of your coworkers. That said their study technique couldn't tell the difference between whether environment attracted the best people, or whether environment caused people to become more productive. (From the point of view of a smart employer there isn't too much difference between those hypotheses.)
4. People accept compensation in many ways, and not all of them involve money. I believe that the better a programmer is, the more apt they are to be willing to trade salary for better work environment, more interesting co-workers, more interesting problems, etc. As a concrete example I have heard anecdotes from many people claiming that people are willing to accept sometimes substantial pay cuts for the privilege of working at Google.
5. Whatever the cause, in the Peopleware study the difference in pay between the top and bottom half of programmers was less than 10%. Furthermore productivity was basically uncorrelated with various factors people look at in the hiring process such as years of experience. That is direct evidence that it is a bad idea to measure productivity by salary.
When you put all of these factors together, you simply can't point to market forces then salaries to discount real performance differences between working programmers.