The more successful vertical companies become at post-training, the less differentiation will come from the raw base model. If Intercom can beat Opus 4.5 on customer service, and Meta’s scaffold can make Sonnet outperform Opus, then the willingness-to-pay premium for the very best frontier model can shrink in production use cases. Of course, the labs’ strongest remaining moat can still be on novel, general-purpose tasks where no domain-specific data exists