I attached a file

OpenGL and BI rendering work here; smoke is not supported in Cycles yet. If you're referring to why you can't see the smoke when rendering, it's because you didn't have a material set up to visualize the volume.

Follow the directions here to get a basic smoke material and manipulate to get the desired effect: http://wiki.blender.org/index.php/Doc:2.6/Manual/Physics/Smoke/Material

r59074 x64 (SVN trunk) on Win7

It is not such a thing to point out here.
I can watch smoke by using smoke modifier.
I point out that a different image is produced when I paint pictures in opengl renderer in the next using internal renderer.

The phenomenon is not caused even if I do a similar thing in the state that chose Cycles renderer with a rendering engine.
It occurs when I carry it out in the state that chose internal renderer with a rendering engine.

Because I took an animation, I attach it.
It is the file of the wmv format.

I understand what you are saying now.

Using the video as a guide, I have reproduced the error, but my result is slightly different. I can also confirm this in 2.68a and in SVN trunk, but not in previous versions.

-- My steps to reproduce the error:
Open blend file and bake the smoke simulation to ensure it is cached.
Render in openGL (it will render as expected).
Ensure the selected render engine is Blender Render and render. It will render the domain as the cube (expected).
Immediately render with openGL again. My result was half-opacity smoke over the icosphere. This is the erroneous render.
Immediately render with openGL again. It will render as expected (identical to the first openGL render).

I am able to reproduce this multiple times without reloading the file. Simply render with Blender Render with F12 again followed by another openGL render. Changing frames between the F12 render and the openGL render does not produce the problem. Following the same steps with Cycles enabled does not reproduce the problem.

I will try to find the commit or time period when this was introduced.

The commit that introduced this behavior is 57377.

I have been able to reproduce in 57377 and later with both MSVC 2008 and MinGW-w64 build systems.

Win7 x64
GTX 560 Ti 2-GB version

I couldn't redo this problem (OS X, NVidia GT 650M), but maybe I don't understand it well.

When you say render, is that a single frame or an animation render?

It's a single frame render.

I'm sorry.
There was an old custom called "OBON" in Japan and was absent for these past three days.
There is a thing to get to me just to have understood that it is not a problem peculiar to my environment.

This occurs by the rendering of the animation of opengl.

Miika: could you check on this?

Seems like domain object has incorrect inverse matrix during the first render. So I suppose not everything gets properly updated between renders.

Fix in svn revision 59446, thanks for the report.