Emission scenarios and the IPCC temperature range
Short answer: [1.4,5.8] includes model and scenario uncertainty. fig 9.15 shows the T range across models and scenarios; fig spm-5 gives you a key to the various scenarios (or see this SRES fig). Of those, I think A1T and B1 correspond to fairly drastic action; IS92a was the old business-as-usual approximation I think. My pet preference would be HadCM3/IS92a as a base state: that gives you 3 oC at 2100. Switching to A1F1 would put the change above 5 oC; down to B1 puts it down to 2-and-a-bit.
Or if that is too short, read on:
The IPCC says The globally averaged surface temperature is projected to increase by 1.4 to 5.8°C (Figure 5d) over the period 1990 to 2100. These results are for the full range of 35 SRES scenarios, based on a number of climate models (here). There is a footnote to this, which says: Complex physically based climate models are the main tool for projecting future climate change. In order to explore the full range of scenarios, these are complemented by simple climate models calibrated to yield an equivalent response in temperature and sea level to complex climate models. These projections are obtained using a simple climate model whose climate sensitivity and ocean heat uptake are calibrated to each of seven complex climate models. The climate sensitivity used in the simple model ranges from 1.7 to 4.2°C, which is comparable to the commonly accepted range of 1.5 to 4.5°C.
To interpret that: firstly, the *climate sensitivity* (equilibrium climate sensitivity refers to the equilibrium change in global mean surface temperature following a doubling of the atmospheric (equivalent) CO2 concentration (glossary) is about 1.5 to 4.5 oC. But T change to 2100 isn't necessarily equilibrium, so you can knock a bit (0.5-1.0 perhaps) off that range for non-equilibrium effects (except probably B1 and A1T (see here), which are pretty flat by 2100). This effectively leaves higher CO2 scenarios (like A1F1) adding about 2 oC for the effects of higher CO2.
Secondly, a technical point: all these scenarios haven't been run through all the GCMs, it would be too expensive. Most GCMs have (or had, when the TAR was written) only done a few scenarios. So a simple model is tuned to the model performance in one or two scenarios, and then the model is effectively used to interpolate to the other scenarios. For this post, that doesn't matter.
So, fig 5d shows that an "average model" has a spread of [2,4.5] for all the 35 SRES scenarios. Whereas the spread for all models all SRES is bigger, at the aforementioned [1.4,5.8]. You could try taking a single model and computing the spread across scenarios; or you could take a single scenario and compute the spread across models. The TAR sez: By 2100, the range in the surface temperature response across the group of climate models run with a given scenario is comparable to the range obtained from a single model run with the different SRES scenarios.
Then fig 9.15 provides the answer you were actually looking for (or nearly, because it does six illustrative SRES scenarios (don't ask me what they mean by illustrative) across 7 models, but that should be enough for a flavour of what goes on). Unsurprisingly, a higher base state sensitivity (GFDL) scales to a stronger response and hence greater range across scenarios.
Caveats and stuff. #1: I've never been terribly interested in the range for different scenarios. Clearly its a valid thing to ask, but given the large (AFAIK) uncertainties in all the emissions, I'm more interested in "what is the T change at (say) 2*CO2" rather than "what is the T change at 2100". #2: All this is taken from the IPCC TAR. I don't think things have changed much since then. #3 I (and most people) end to talk in terms of CO2 for simplicity; the SRES scenarios include other GHG's, and sulphates, and stuff too.