37. Kh3 Qg5 threatening mayhem on g4. Best is 38. Qg1 to protect g4; then ...Re7 gets rid of the knight, one way or another, and g4 goes bang anyway.

37. Kg3 Qd6+; Kh3 or Kh4 Qh6 threatening mayhem on the h-file.

37. f6 Rf7 followed by hxg4 and g3+ and the WK has nowhere to hide. If, say, 38. Qf1 hxg4; Kg3 to stop this then Qd6+; Kh4 g3+ and White is toast. In some variations B can just play ...Rxf6 instead of ...hxg4, and again the WK ends up with nowhere to hide.

(I'm a patzer, or more precisely a rusty club-level player; the above is the result of some experimenting with a strong computer program. I may have missed important things.)

36...Qd8+ only looks better to a computer. The idea is 37. f6 hxg4 38.Nxg4 gxf6 threatening ...Rh7+ or 37. Kg3 Qd6+ 38. Kh4 Qh6!

But Anand's 36...g5+ is much easier for a human to play. It's easier to see White's reasonable responses and how to deal with them.

Anand never made a move that let go of his winning advantage, and all the (mild) criticisms of his play are based on moves that computers think win a little faster; but a human should play to win as reliably as possible, and that's what he did.

I am astonished though by Topalov's mistakes capturing the f and e pawns. They were inexplicable, because they're not really tactical errors at all; they simply show terrible positional misunderstanding.