Well, then I would say the largest limit is the driver. I don't know anything about the driver, but read on.
The OPL3 chip has 36 operators, that can be divided in many ways into channels.
Simplest mode just has 18 channels with 2 FM operators each. The 2 operator channels can use two different sound producing algorithms, just summed together or it can produce FM sound with a modulator and a carrier.
It is possible to reserve 6 of the operators for drum mode with five drum sounds (bass drum, snare drum, tom tom, cymbal and high hat).
But that will take away three 2-operator channels.
Also you can have up to 6 channels of 4 operator FM sounds. The 4-op channels can use 4 different sound producing algorithms.
This leaves remaining 6 operators into three 2-operator channels.
You can select if a channel is heard on left, right or both speakers, but there is no panning mechanism in the chip.
Each operator has an ADSR (attack-decay-sustain-release) envelope, volume, frequency, note trigger on/off, etc and thus changing the parameters of modulator changes the timbre and how the timbre changes over time duration of the sound, while changing the parameters of carrier changes how the sound volume changes over time duration of the sound.
Now, here comes in the driver. MIDI information is nothing more than note on/off triggering information, selecting instruments and volume and pitch bend. It is up to the driver how good OPL3 chip parameters it uses to recreate a sound that resembles a piano or guitar or whatever. Also it is up to the driver to use any tricks to make the sound even better, like updating the chip parameters tens or hundreds of times per second to improve the sound to resemble a piano or guitar or whatever better. Not many drivers do this, but some game audio systems do.
But basically, the reason while FM sounds are very dull and boring is the fact that most often the instrument parameters are just loaded to chip at song start and only frequency information is updated and notes are just triggered on and off. This also consumes least amount of CPU time.
A driver can do some stuff, like for example you could pan a sound by using two channels, one for left and one for right, with same or a bit different parameters, and just controlling the volumes of the channels individually so that the sound can be heard say 25% left and 75% right.
Another thing is, is the driver for an OPL3 or OPL2.. OPL2 requires much slower delays between writing to the chip, and only has 18 operators and it only has 9 melodic 2-op channels or 6 melodic 2-op channels plus the 5 percussive channels, single mono output. So because OPL2 is slower to write, some of the tricks may not be used as it would consume too much CPU horsepower to update chip registers tens or hundreds of times per second. (Relatively speaking of course. I approximate that OPL3 chips were available when people had 386s and above, while OPL2 chips were used before and up to 386s)
