As many companies cut AI costs, Shopify bars engineers from using non-frontier models and startups like Olive and Avoca AI prioritize accuracy over token costs
Despite steep and rising price tags, some companies heavily favor the more powerful ‘frontier’ AI systems over cheaper alternatives
Context & Ripple Effects
The coverage has framed AI procurement as a split between companies lowering inference bills with cheaper models and a smaller group willing to pay for stronger output. Shopify’s policy was reported in the earlier account of its frontier-model-only approach, making this a clearer example of the quality-first side of that divide.
That split follows reports that rising AI costs are pushing some buyers toward cheaper options, including Chinese models, putting pressure on frontier-model pricing. Olive and Avoca AI place themselves with buyers for whom accuracy is judged more important than token savings.
First-order effects
- Shopify engineers lose the option to choose non-frontier models, standardizing internal AI use around the most capable systems rather than the lowest-cost ones.
- Olive and Avoca AI accept higher token spending where they believe better model accuracy improves their product’s output.
Second-order effects
- The gap between low-cost and quality-first buyers becomes more explicit: providers compete not only on token price, but on whether their models can justify use in accuracy-sensitive workflows.
- Companies adopting similar policies will need to evaluate AI spending against useful output and error reduction, rather than treating token cost as the sole procurement metric.
Third-order effects
- If this divide persists, AI procurement could bifurcate into cost-optimized workloads and frontier-dependent workflows, with model selection increasingly tied to the consequence of mistakes.
- The longer-term advantage may accrue to providers that can demonstrate superior outcomes at specific tasks, while efficient smaller-model approaches remain viable for less demanding work.
The trend: AI adoption is moving from broad experimentation toward workload-specific procurement that weighs cost per useful result against the value of higher accuracy.