As AI models keep changing fast, is it still worth building your own tools — or just buying them?

This explores how fast-changing AI models shift the cost of building your own AI tools compared with buying them from vendors. The clearest evidence is a reversal. In a16z's 2025 survey of 100 enterprise CIOs, companies were moving away from custom AI apps and toward third-party products. The reason they gave wasn't that their in-house tools were bad. It was that every new model release meant more re-optimization and maintenance, until the internal tool stopped giving them any edge Are enterprises abandoning custom AI apps for third-party solutions?. Rapid model change works like a tax on whatever you build, and the tax recurs.

The useful question is which layer that tax lands on. Research on self-improving agents splits improvement into two loops. The slow loop retrains the model's weights. The fast loop updates the prompts, memory and tools around the model. Most recent progress has come from the fast loop because those changes are cheaper and easy to undo Do self-improving agents really split into two distinct loops?. One harness study backs this up from the practical side. An optimized execution setup, built around models that were never retrained, raised Terminal-Bench scores across several models, and the same runbook carried over to newer models without modification Can execution harnesses lift model performance without retuning weights?. So build-versus-buy may be the wrong split. The better question is what to own. Work that is tightly tied to one model goes stale quickly. Workflow, tooling and execution logic can outlast several model generations.

There is a catch. Another study finds that the benefit of a harness follows an inverted U across model tiers. Mid-tier models gain the most. Weak models fail to use the harness at all, and the strongest models are worse at following its instructions faithfully Do stronger models always evolve harnesses better?. A scaffold that pays off today could matter less after you upgrade. The economics also move as smaller models improve. A 35B model trained on execution-heavy data now sits at the low-cost end of the cost-performance frontier for multi-step work Does model efficiency matter more than peak capability for real work?. That makes running and tuning your own mid-sized model look more viable than it did when only the largest models could do the job.

The gains from building are uneven. Firms with more AI exposure replace contracted labor with AI faster and more cheaply than less-exposed firms, which suggests that in-house AI capability gets cheaper the more of it you have Do firms substitute labor for AI at different rates?. For a firm that already has that capability, building can keep getting cheaper. For a firm without it, the maintenance tax a16z describes probably dominates. One warning applies to both: don't judge a model or vendor by benchmark wins. Agents that top benchmarks still fail at long, realistic professional workflows Why do agent benchmarks not predict real economic value?.

A direct caveat: the corpus has little material on build-versus-buy as such, and the a16z survey is the only source that addresses it head-on. The rest is lateral evidence about which kinds of investment survive model churn. Taken together, it suggests that what lasts through model churn is the layer around the model, while work tied to one particular model loses value with each release.