While I have seen many papers in which Gaussian is used in calculations, I have not seen any paper in recent times where new methods are implemented in Gaussian. The latest paper I could find concerned the EDIIS method which was published nearly twenty years ago. Development of methods seems to happen in Q-Chem, Psi4 and TURBOMOLE.
Is Gaussian too old, the code too deprecated or are there copyright issues?
"Is Gaussian too old, the code too deprecated or are there copyright issues?"
You mentioned "copyright issues" but papers that are publicly available in academic journals are not subject to the type of copyright laws that would prevent Gaussian from being allowed to implement them.
You suggested that the code is too deprecated but if you look here then click on "Collaborators", you'll see a lot of academic researchers who appear as co-authors of at least one version of Gaussian, and still actively develop new methods, often using a developer version of Gaussian as a starting sandbox.
The other suggestion you made was that GAUSSIAN is too old, which is also not the case. MOLPRO development started in the 1960s, yet several new features were implemented in the following new versions: 2021.2, 2021.1, 2020.2, 2020.1, 2019.2, 2018.1, etc. These include, the AVAS method which was first in 2017, and PNO-LCCSD(T) and PNO-LCCSD(T)-F12 and PNO-CASSCF and PNO-CASPT2 (which likely weren't available anywhere before 2015, though they are similar in spirit to the DLPNO methods in ORCA, which were also implemented for the first time much more recently than the nearly 20-years ago you mentioned for the development of EDIIS).
So if it's true that MOLPRO, Q-Chem, Psi4, and TURBOMOLE are still implementing newly developed methods, substantially more than Gaussian, and you want to know why this might be, then read on:
Psi4 is open-source, so I'm not surprised that it's still being actively developed and contains implementations of newly developed methods.
None of the other software packages that you mentioned are open-source, so we can talk about those now. The newly developed methods that keep getting implemented in MOLPRO, Q-Chem, and TURBOMOLE are largely the developments of active researchers who are developers of those same software packages. It was asked here whether any non-free software packages have implemented any methods which are not available in any free software and also have been used in publications that are not by the developers of that method: What significant matter modelling methods are implemented in commercial software, for which there is no freeware alternative? No answers have been given to that question yet! Specifically, I mentioned AVAS in MOLPRO, for which the original paper that I cited there had Gerald Knizia as the last-listed author, and Gerald happens to have also been one of the major contributors to MOLPRO (notice he is the [first author of the 2012 version, apart from the Werner an Knowles who happen to be the two main authors of the package). I also mentioned the PNO methods in MOLPRO, which according to the bibliography section of this page were implemented based on 13 papers, all with Werner a co-author.
Similarly, the current Q-Chem manual says that version 5.4 is the most recent, and the new features in 5.4 include methods like incremental FCI based on Paul Zimmerman's papers, and Paul happens to be a co-author on the 2015 Q-Chem paper. All members of the Q-Chem Board of Directors except for Hilary Pople, are highly active researchers at universities.
TURBOMOLE is similar, with the company being run by Wim Klopper (Karlsruhe Institute of Technology), David Tew (Oxford University), Christof Hättig (Ruhr-Universität Bochum), etc.
The last three paragraphs are in contrast to Gaussian, where if you go here again but this time don't click anything, you'll see that Gaussian is run by employees who are full-time staff members of the company, and researchers who are running larger research groups of their own tend to be listed as "Collaborators". This is in sharp contrast to what I described for MOLRPO, Q-Chem and TURBOMOLE, and if more academic research development is getting implemented into the latter three software packages than in Gaussian, it may be largely due to this difference in the extent to which they've become "commercialized" (they're all commercial software packages, but they all began with researchers from educational institutions and the extent to which Gaussian's culture and main employees transitioned from educational-type environments to their newer corporate-type environment, seems to differ greatly considering that MOLPRO, Q-Chem and TURBOMOLE are all run by people still at universities whereas Gaussian is run by their own full-time employees while researchers at educational institutions are relegated to the label of "collaborators").
Also: Gaussian has historically probably been the most popular of all of these software packages, in terms of the sheer number of users that have used it. When a user-base is as big as theirs, it's not always as easy to make changes to the code with consequences as mild as if the user-base is smaller.
Just to add 2 little cents regarding new methods in Gaussian - the extract from the Gaussian site (www.gaussian.com) about new methods in Gaussian 16:
New Methods Many DFT functionals have been added to Gaussian since the initial release of G09, including APFD [Austin12], functionals from the Truhlar group (most recently MN15 and MN15L [Yu16]) and PW6B95 & PW6B95D3 [Zhao05a]. Additional double-hybrid methods: DSDPBEP86 [Kozuch11], PBE0DH, PBEQIDH [Bremond11,Bremond14]. Empirical dispersion for a variety of functionals, using the schemes of Grimme (GD2, GD3, GD3BJ) [Grimme06, Grimme10, Grimme11] and others. The PM7 semi-empirical method, both in the original formulation [Stewart13], and with modifications for continuous potential energy surfaces [Throssel17p].
