The AI Model Market Did Not Consolidate, It Fragmented

Research on pricing elasticity confirms what practitioners observe: for professional workloads, AI usage is largely price-inelastic. A 10% drop in token costs produces roughly 0.5% to 0.7% increase in consumption. The people using these models for real work are not shopping around based on cents per million tokens. They are locked into whatever model fits their workflow. Developers call this the glass slipper effect: once a model works well within a specific toolchain and habit, switching costs are real even when the financial friction is low.

The Cost Collapse Is Real, Just Not Where You Looked

The dramatic price reductions in AI inference are genuine. Gemini Flash at roughly /bin/bash.10 per million input tokens represents a remarkable comedown from where the market sat two years prior. This is not marginal improvement. It is an order-of-magnitude shift that has made AI economically viable for high-volume, low-stakes tasks at scale.

But the cost collapse is concentrated in routine, commodity work: summarization, classification, content generation at scale, first-draft extraction. These are valuable markets. They are also increasingly undifferentiated. When anyone can offer sub-cent per query inference, the competitive moat is not in the model layer. It is in the application layer, in distribution, workflow integration, and vertical fit.

The irony is that the price collapse, while real, has barely moved professional adoption curves. The 0.5-0.7% usage bump from a 10% price cut tells you everything: the professionals who matter most are not particularly responsive to cost. They are responsive to capability and fit.

The Surprise Surge

Among the more striking data points in recent months: DeepSeek's rise on the OpenRouter aggregation platform. From a standing start, it captured roughly 30% of OpenRouter traffic, a remarkable redistribution that happened faster than most observers anticipated.

DeepSeek is not the best model on every benchmark. It is not the most expensive. It is a capable, well-positioned open-weight model with a price-performance profile that certain workflows find genuinely compelling. The lesson is not that DeepSeek won. The lesson is that a market with multiple viable options creates space for specialized competitors to find their audience quickly, without needing to displace the incumbents across the board.

The Frontier Illusion

ARC-AGI-3 scores offer a useful corrective to the narrative of relentless frontier progress. Every leading model, without exception, scores below 1% on the variant testing radical generalization. This is not a cherry-picked failure mode. It is a signal that the benchmark is measuring something the current generation of models cannot do, regardless of scale.

This matters for market structure. If the frontier is not meaningfully advancing on certain task categories, then the differentiation available to competitors is not just price and efficiency. It is the quality of specialization in the tasks where progress is possible. A model that is exceptional at a narrow set of tasks is more valuable to a specific user than a generalist that is merely good at everything.

What Fragmentation Actually Looks Like

The market is not converging toward a single best model. It is splitting along the lines of genuine user need. Claude in programming. DeepSeek in cost-sensitive, open-weight scenarios. Gemini Flash in high-volume commodity inference. GPT-4o maintaining broad consumer reach while its internal efficiency has been optimized in ways that may have affected output characteristics over time.

None of these positions are permanent. All of them are sticky in the near term. The glass slipper effect cuts both ways: incumbents benefit from lock-in, but entrants can exploit it by finding unguarded niches. The next significant market share shift will not come from a model that is simply better across the board. It will come from a model that fits a specific workflow better than anything currently serving it.

The consolidation story was always a projection from past technology cycles onto a market that turned out to behave differently. Infrastructure may favor scale, but user behavior, workflow integration, and task-specific fit create durable fragmentation that scale alone cannot overcome. The AI model market did not consolidate. It fragmented along the grain of how people actually work.

Apr 21, 2026