From the included whatsnew.txt file:

1) Faster FFTs for most SSE2-capable CPU architectures.
2) More FFT lengths supported.
3) Some of the new, faster FFTs don't let us do as much error checking on each iteration.
If you don't want to lose this error checking, choose "SUM(INPUTS) error checking" from
the Options menu. This will slow down your tests by about 2-5%.
4) When a possible hardware error occurs during a test, the error message would
likely scroll off the screen unnoticed. Now, all future progress updates to
the screen will be followed by a message containing a count of the errors that
have occurred during the test.
5) Worker Windows now stagger their starts. This improves timings for some
CPUs and operating systems. See undoc.txt to control this feature.

Known bugs and fixes:

1) The length 7K real FFT crashes on Core i3/i5/i7/i9 when running 32-bit Prime95 or 32-bit mprime. Fixed in next release.
2) Benchmarking on a multi-core or hyperthreaded machine resulted in several spurious "Setting affinity to run help thread..." messages. Fixed in 26.3.
3) The length 3072K (and some others) FFT crash on Northwood and Willamette Pentium 4. Fixed in 26.3.
4) The new feature that continuously outputs hardware error counts took multiple lines. Version 26.3 will have several undoc.txt options to control the amount of verbiage.
5) In rare situations, an insufficiently large FFT length was chosen. Generally, you would only notice this with larger bases. An example is 10024*603^153-1. Fixed in 26.3.
6) Worker-specific memory settings were not used in calculating the optimal P-1 bounds. Fixed in 26.3.
7) On 64-bit Ubuntu 10.x machines, trial factoring would find factors but would improperly format them for writing to results.txt and for sending to PrimeNet. Fixed in 26.3.
8) On 64-bit Ubuntu 10.x machines, N/A in worktodo.txt caused parse errors. Fixed in 26.3.

1) The length 160K x87 FFT crashes. Fixed in 26.3 build 2.
2) The length 4M FFT crashes on Northwood and Willamette Pentium 4. Fixed in 26.3 build 3.
3) On a brand new install, the default on hyperthreaded machines was 2 threads per worker. This was slower for most users. In the next release, the default will be 1 thread per worker.
4) Testing 522*928^24576-1 in 64-bit executables crashes. Fixed in 26.4.
5) You cannot edit the affinity settings from the Linux menus. Fixed in 26.4.
6) Running a multi-threaded torture test on AMD machines failed. Fixed in 26.4.
7) Setting a specific FFT length in worktodo.ini sometimes selects the next larger FFT length. This would most commonly occur when an exponent being tested is near the FFT crossover point and a specific FFT length is automatically chosen based on the roundoff error in the first 1000 iterations. Fixed in 26.4.
8) Multi-line additional information from the Primenet server was not parse properly. Only the first line was parsed. Thus, when successfully completing a double-check only the successful completion was written to prime.log. The cpu credit amount was not written to prime.log. Fixed in 26.4.
9) When doing PRP on non-base-2 numbers the time estimate in Test/Status was off by a factor of log2(base). Fixed in 26.4.
10) On a 32-bit Mac, some FFT implementations were missing. Fixed in next release.