New in 3.8 (since 3.6)

- many small improvements and tuning since 3.6
- experimental Monte Carlo mode (9x9 only)
- support for tiny boards
- GPL v3

New in 3.6 (since 3.4)

- stronger than 3.4
- many small improvements
- GNU Go can now resign games
- emacs mode can display the board graphically using xpms
- basic knowledge of how to break out of mirror go

New in 3.4 (since 3.2)

- stronger than 3.2
- new connection and semeai modules use lookahead
- new break-in code
- improvements to many parts of the program

New in 3.2 (since 3.0)

- stronger than 3.0
- uses less RAM than 3.0
- 1-dimensional board
- experimental dynamic connection analysis
- experimental reading semeai module
- new influence function
- stronger and more agressive
- reads to find combinations
- revised Zobrist hashing
- new html views of the regressions, and many more tests

New in 3.0 (since 2.6)

- stronger than 2.6
- new move generation scheme
- new influence function
- more accurate reading
- board information maintained incrementally during reading
- new "owl" and "life" modules for accurate life and death analysis
- persistent caching of reading and owl results for speed
- revised semeai module
- experimental Deterministic Finite State Automaton (DFA) pattern matcher
- new debugging tools
- level option

New in 2.6 (since 2.4)

- stronger than 2.4
- more portable code
- Texinfo documentation
- Emacs mode

New in 2.4 (since 2.0)

Enhancements for the user:

- stronger: able to give GNU Go 2.0 a 5 stone handicap
  - life and death evaluation drastically improved
  - more efficient and accurate reading
  - (small) joseki database
  - takes influence and territory into account
- Ascii interface as an alternative to CGoban
- uses GNU configure
- undo supported
- man page

Enhancements for the developer:

- expanded pattern database
- autohelpers for patterns
- joseki library in Smart Go Format
- fuseki module
- backfilling and numerous other improvements to reading code
- algorithms for estimating territory and influence
- eye_finder module uses a static algorithm for life and death
- eyeshape database
- connection database
- reading code uses Zobrist hashing and other speedups
- greater modularity
- documentation of key algorithms
- expanded support for Smart Go Format
- various debugging tools