cubic 2.0

Over the past few months, we've been completely rebuilding cubic's AI review engine.

Today we're excited to announce cubic 2.0, the most accurate AI code reviewer available.

The numbers

Compared to three months ago:

• 3x more actionable: 20% → 60% of comments addressed in follow-up commits

• 2x faster: median time to review a PR was halved

• 40% better signal: upvote ratio went from 1.05 to 1.47

How we did it

Accuracy

AI review tools get noisy when they don't understand your repo. The diff might look wrong, but it's actually how you do things. Or the diff looks fine, but it breaks something specific to your setup.

We made a lot of changes here. Some examples:

• Repo context: cubic now automatically reads READMEs, contributing guides, and context files in your repo to understand how things are supposed to work before it comments

• Live documentation: cubic smartly reads the correct docs for the lib version you’re using

• Better tooling: improved the tools cubic has access to crawl and fetch the relevant code and files it needs

• Filtering: all of this context feeds into filtering out bad flags before they get posted

Speed

PRs are iterative. You push, get feedback, fix something, push again. Before, cubic would re-clone and re-analyze everything each time.

Now we cache codebases for a short window. Back-to-back pushes on the same repo skip the clone step entirely. This made reviews a lot faster, especially on bigger repos.

cubic vs the competition

We compared cubic against other AI code review tools on repos running both.

CodeRabbit

On repos running both cubic and CodeRabbit, cubic flags 50% more unique issues that users end up addressing. These are all bugs that users fix that CodeRabbit did not flag at all.

Additionally, 80% of comments that CodeRabbit posts that cubic doesn't end up not getting addressed by the user.

Cursor

On repos running both cubic and Cursor, cubic flags 2x more unique issues that users end up addressing. These are all bugs that users fix that Cursor did not flag at all.

The data shows cubic catches more of what matters.

Try cubic 2.0 on your next PR → Get started