It is widely believed that stereopsis is only useful at near distances and there are few stereoscopic studies at distances beyond a few metres. Allison, Gillam and Vecellio (2009) found stereoscopic depth estimates to be much more accurate than monocular up to 9 metres. In the present studies we extended these studies to 40 metres using a disused railway tunnel. Bluetooth-controlled pairs of LEDs were separated in depth and viewed either in complete darkness or with the space between the observer and the nearest LED illuminated to introduce perspective cues to distance (but not to the depth between the LEDs). Near LEDs were at a depth of either 20 or 40 metres, and the far LEDs up to 7.76 metres or 31.03 metres further, respectively (an equivalent disparity range was tested at both distances). Naive and non-naive observers verbally reported the apparent depth interval (in metres) between the LEDs in each pair. In the dark, observers greatly underestimated real depth intervals, with maximum estimated depths of 1.0 and 1.4 metres, for 20 and 40 metre distances. When perspective cues to distance were available, maximum depth estimates for these distances were 3.3 and 4.4 metres. We conclude that for the large absolute distances studied, stereoscopic depth intervals, while varying with disparity and distance and showing a clear effect of scaling by perspective cues, are greatly underestimated.