Dear Christian,

This sort of question has arisen repeatedly in the recent past, therefore I will try to give a general and (hopefully) simple summary of the changes that come with DCM10.

DCM10 for fMRI (which is part of the SPM8 release since version 4010 in July 2010) includes numerous additional features and various changes compared to "classical" DCM (now referred to as DCM8).

In DCM10 for fMRI, you can now specify a model as
- bilinear or nonlinear
- single-state or two-state
- deterministic or stochastic

You can choose any combination of these options. This 2x2x2 factorial structure gives you 8 options to specify the general nature of the DCM.

Importantly, Karl rewrote the general DCM code to make it universally suitable for all of the above combinations.  Some of these changes have a noticeable impact on the specification and implementation of a classical (bilinear, single-state, deterministic) DCM.  A non-exhaustive list of relevant changes includes:

1. Inputs are mean-centred.
2. The prior variance of endogenous and bilinear coupling parameters no longer depends on the number of areas, but is fixed.
3. A simplified version of the hemodynamic model is used.
4. Self-connections are estimated for each area separately.
5. Various components of the optimisation procedure have been adapted.

Note that these changes mean that DCM10 and DCM8 implement different models and can yield different parameter estimates, even when you are using a classical (bilinear, single-state, deterministic) DCM.

Furthermore, the mean-centering of inputs means that the interpretation of endogenous connection strengths (A matrix) changes. This is because the expansion point of the Taylor series is now the average of the input (u), therefore A no longer represents "baseline" connectivity (in the absence of perturbation), but the average connectivity (i.e., under average perturbation); this perspective is particularly useful for stochastic DCMs.  As before, B still designates additive increases/decreases in coupling.

As some users may prefer the traditional interpretation for their specific applications, the next release of SPM will give users the option whether they want to mean-correct inputs or not.  This allows you to interpret the endogenous couplings from your preferred perspective.

When publishing papers, you should state whether you are using DCM10 or classical DCM (now referred to as DCM8).  If you are not sure, check the MATLAB command window when using DCM: DCM10 reveals its presence by printing a message there.

I hope this summary is useful. 
With my best wishes,
Klaas