Predicting AI timelines has a poor track record, so it is more useful to look at what is changing structurally than to guess at capability milestones.
Cost per unit of capability continues to fall substantially. This has a practical consequence: applications that are uneconomic today may be viable within a planning cycle. Architecting so the model is a swappable component rather than a hard dependency preserves that option.
Smaller specialised models are closing the gap with frontier models on narrow tasks. For well-defined problems — classification, extraction, structured summarisation — a small tuned model is often cheaper, faster, and easier to run in your own environment than a general-purpose API.
Context windows have grown dramatically, which reduces but does not eliminate the need for retrieval. Attention over very long context still degrades, and sending large contexts is expensive. Good retrieval remains cheaper and more accurate than brute-force inclusion.
Multimodal capability is becoming the default rather than a differentiator. Systems that only accept text will look increasingly limited as documents, images, and audio become routine inputs.
The genuine bottleneck for most organisations is not model capability. It is data readiness, evaluation infrastructure, and integration with existing systems. Teams that invest there will adopt each capability improvement faster than teams waiting for a model that removes the need.
Regulation is tightening around transparency, data provenance, and high-risk applications. Building with auditability — logging what was retrieved, what was sent, and what came back — is likely to be a compliance requirement rather than an engineering nicety.