It's not that the transistor is charging and discharging things. That works for constant currents and fixed voltages and slew rates, but that has to do with a certain nonlinearity. Optos tend to be more linear than that.
A good model of an optotransistor is a variable current sink (in parallel with a diode so it doesn't pull below 0V) in parallel with a capacitor, which is in turn supplied by the external resistor and supply.
The frequency response of this model is simply Vo = Iin * (R || Zc), so it rolls off (-3dB) at the frequency when R = Xc. The gain is Vo = Iin * R.
If you make R smaller, the cutoff frequency rises, and the gain falls proportionately. You get less than 1V of signal for a 100 ohm load, but at least you get transitions in the single microseconds range.
What if you want more? You can cheat by using external amplifiers. Here's an example:
The 4N35 collector 'tugs' on a PNP emitter, which has a very small load resistor pulling up to the supply. In fact, the impedance seen by the 4N35 will be even less than 51 ohms, more like 2 ohms as shown (the small signal equivalent emitter resistance of the PNP). This is called "folded cascode", 'folded' because the input current is pulling down, yet the output (the PNP collector) is pulling up, and 'cascode' meaning a collector driving another emitter. (The three other transistors in this circuit only serve to supply bias voltage / current, or buffer the output, though they are nonetheless important to operation. One cheat: the 4N35 transistor is itself being used as an amplifier; the circuit works just as well with something like 6N138 though.)
Anyway, with the R feeding the 4N35 being as small as possible, the speed is as high as can be, under a microsecond. The gain is also quite linear. This, I believe, is about as close to ideal / theoretical as a 4N35 can get; ultimately, performance is limited by the structure of the phototransistor itself, which is optimized for optical gain, at expense to speed, breakdown voltage and other design variables.
For general purposes, I would much rather suggest a 6N138 (faster, simple to use, cheap), or one of many logic type isolators (optical or inductive) that are available.
Tim