According to the classical model developed by Evans and co-workers on the double torsion test [(1972) J Mater Sci 7:1137 and (1973) J Testing Eval 1:264], the stress intensity factor is independent of the crack length. Recent applications and analysis question this independency (Chevalier et al (1996) Cer Inter 22:171, Ciccotti et al (2000) Inter J Rock Mech Min Sci 37:1103). This work consists of using samples with different lengths of a typical brittle material (a soda-lime glass) in order to discuss on the validity of the different equations proposed to analyse the DT technique. Experimental compliance tests always showed linear variations with crack length. Successive relaxation tests revealed, however, a clear dependency of the stress intensity factor on crack length. This dependency, observed through the non reproducibility of the V-K-I diagrams, is reduced as the sample length increases. The corrections proposed by Chevalier and Ciccotti on Evans model revealed that their applications remain limited to the sample and the loading configurations used by the authors. The application of Evans model without correction is conditioned by the use of sufficiently long samples and advanced crack lengths.