I kinda like Unity, but my current project using it, heavily uses file dialoques, which unhappily are tending to crash it. And that isn´t really a lot of fun, as this means restarting Unity every 15mins or so.

Are you working on an iphone app/game?

Nope, at least not only.

I btw found out that Unity renders each light in a new pass, which should actually be quite slow...

Unity 3 uses deferred rendering, so this behaviour will be history then.
The Beta is already available.

Deferred rendering is not everything. It is very hard to implement in a flexible way, if that is even possible at all. But well, I was told that the shader system was improved a lot, so I am looking forward to the free edition of 3.0 And I could imagine that the free Unity 3.0 does not support deferred rendering, at least not if they don´t make post processing effects available in the free edition .

But they advertise that hundreds of lights are possible with the new deferred rendering approach. Since I have license of the pro edition now, I will check this extensively.

that is the point of the deferred shading -hundreds of lights- but it comes with the expense of the losing flexibity.

What flexibility do you have in mind?

As far as I saw all big (and expensive) engines (including CryEngine 3) use deferred rendering, so there must be good reasons for using it.

flexibity of rendering process and using shaders. It does not affect the artist in anyway, only the shader programmer.

but as Nils said, it's "very hard" to implement "wihtout" losing flexibity. Crytek makes hard things real... I used cryengine3 for a week about 7-8 months ago, integrating custom shaders was dead-easy almost didnot required an extra step compared to forward rendering. It is also up to the user to use deferred shading in cryengine3.

Well, one for example have to deal with transparency, refraction and reflection, as you only have some textures for deferred rendering, which for example contain the depth or the normals of the scene. And you can only have the normal of of one surface at one pixel and not the one of several surfaces. Because of this, deferred rendering is often not used for such surfaces, which means that you will always need a second shader for everything. There are some alternative solutions, but they have drawbacks as well. The whole thing is very pixelshader intensive and thus won´t work at acceptable speed on mobile devices, so they for sure don´t remove the possibility for forward rendering, which would be stupid anyways, as one may wants to transform vertices or put some extra information into an extra texture or transforming the normal and diffuse texture for parallax mapping.
The problem is, that the whole pipeline is kinda complex and everything has to fit together. While with a forward rendering approach, you just have one shader per object and they just don´t influence each other.
The whole stuff will be perfectly fine, when sticking to the builtin shaders, but something custom probably doesn´t get easier.