How should SaaS pricing evolve when GenAI introduces unpredictable compute costs and outcome variability?

SaaS pricing must adapt to GenAI’s unpredictable compute costs by shifting toward usage-based, hybrid, or outcome-driven models that align costs with real value. Transparency in AI resource consumption, commitment discounts, and smarter cost controls will be crucial to prevent overruns. For IT leaders, this means closer monitoring, governance, and strategic planning to balance innovation with financial predictability.

I agree that companies should not charge clients to be beta testers, but that is a part of many business models. In the case of AI functionality, the provider will use the client’s interaction with the service, or sometime client data, to improve the offering.
Reducing the cost of a new offering can work to incentivize client to take the journey with the provider. Charging full price for erratic AI portions of a service will probably result in losing clients.
Directly answering the question:
$0 for unfinished
Reduced price for beta
Higher price for stable product with added functionality (that is useful to the client, as opposed to useless new functions) 

My first thought is that the GenAI option should be a premium offering.  But then I considered that this would be a terrible idea.  The best vendors will necessarily fold AI into everything they do or be left in the dust.  I'd imagine that SaaS vendors need to build in capability to manage licensing and costs like the hyper-scalers do with billing alerts and thresholds.

NONE!  If your GenAI system is introducing unpredictable costs and outcome variability, its not ready for production and should not be an active portion of proper pricing or design.
  Poorly executed GenAI should have NO impact to SaaS pricing. SaaS (or ANY as-a-Service) as pricing should only be designed and billed with accurate input, models, logic and output. 
  Your AI integration plans should have financially accounted for adequate development, data integrity and planned to run GenAI model in parallel with existing pricing until not only accurate, but provides more value than existing methods... this means Zero unexpected/unpredictable costs.