
A dramatic cost structure is emerging in AI: “good enough” models are collapsing toward free while frontier capabilities maintain premium pricing. This barbell pattern has profound implications for how companies should think about AI investment.
The Cost Collapse
GPT-4 level performance dropped from $37.50 to $0.14 per million tokens—a 99.7% reduction in under three years. Multiple models now achieve human PhD-level benchmarks for under $1 per million tokens. DeepSeek r1’s January 2025 release at ~$1/million tokens forced industry-wide price cuts.
This isn’t gradual commoditization. It’s a cliff.
The Frontier Premium
Yet cutting-edge models like Claude Opus 4.5 maintain premium pricing. Limited supply, high demand, and genuine capability differentiation preserve pricing power at the frontier. The pattern mirrors pharmaceuticals: generics collapse in price while novel treatments command premiums.
The Strategic Framework
Apply business model thinking: this barbell creates two distinct strategies.
For commodity tasks: Last year’s best is nearly free. Routine classification, basic generation, standard extraction—these should cost approaching zero. Any company paying premium prices for commodity AI tasks is misallocating resources.
For frontier needs: Today’s best still commands premium. Complex reasoning, novel problem-solving, cutting-edge capability—expect to pay. The question is whether your use case genuinely requires frontier or whether “good enough” suffices.
The Second-Order Effect
As second-order thinking suggests, examine downstream implications. If commodity AI approaches free, competitive advantage shifts entirely to frontier access and implementation expertise. Companies optimizing for cheap AI may find they’ve optimized for table stakes while competitors captured differentiation through frontier capabilities.
The bottom line: AI costs are collapsing behind the frontier, not at it. Know which side of the barbell your use case sits on—and price accordingly.









