How often are the underlying AI models used by Ramp's Stack platform updated, and is output consistency part of the selection criteria?

The team is constantly evaluating models and updating benchmarking. As an example, they are currently using Claude 4.6 rather than 4.7 because their benchmarking shows 4.6 performs better for accounting work, even though 4.7 is available. They have relationships with the major AI labs and get early access to models (for example, they had access to a newer Claude model, referred to as 'Fable,' that was later pulled back from the market). Consistency and accuracy are part of what's evaluated, and the goal is that as models change in the background, users should see improved accuracy, lower token usage, and faster output—never a regression from what they're used to.