The answer to your question depends on many factors:
- Technology node. The smaller the transistors get, it's harder to model them. There are effects that are included in models, such as Drain Induced Barrier Lowering (DIBL) - effect of drain voltage to Vth. But there are effects that are NOT included in models, or at best as estimate. For example there is Well Proximity Effect (WPE) - which should model the behaviour of transistor when it is close to well (such as n-well) edge. TSMC has an option in 90 nm technology setup to include this effect or not. I can't tell you (because I don't know) how accurate the models are for WPE. Once more, it is clear that there is no tool that can compensate for good layout.
- Process target market. If the process is intended for pure digital circuits, don't expect accurate models for weak inversion or sub-treshold. I have designed some circuits in technology which is intended for analog/mixed-signal applications, and we have measured fair agreement with simulation. There were some differences in extreme conditions (such as higher than normal voltage or temperature), but you can have a good estimate of circuit behaviour. Some colleagues of mine have fabricated sub-treshold circuits and they have measured significant discrepancy with simulation. The errors are mostly related to leakage currents.
- Circuit topology. If you are designing an op-amp, it's not that important if the gain is 1e6 or 3e6. The exact amount of gain might be very important if you're designing a 32 KHz low power oscillator. I have designed both of these in 0.18 um technology, and many transistors were in weak inversion. In general, if your circuit has a feedback and you only have to provide some minimum gain, then it should be OK. Even if it seems that the circuit can't be made with feedback (32 kHz oscillator), there might be a way to make it exhibit a self-tuning behaviour (in osc. case it will include a Bessel function behaviour, but it is still a feedback

- There is a big fuss about EKV models, which should have smooth transition from region to region of operation. I can't tell if this is an academic "duck" (i.e. someone needed a PhD) or a real thing, but Spectre supports EKV models.
I hope this will help you. If you have something specific in mind, please send your idea, maybe someone can help you.