The share of AI use accounted for by cost-effective AI models is growing steadily.

Tomasz Tunguz (토마즈 턴구즈), founder and general partner of venture capital firm Theory Ventures, highlighted in a recent post he shared on X, formerly Twitter, that while the performance of the latest AI models keeps improving, actual usage is concentrating not on the highest-performing models but on cheaper models with slightly weaker performance.

He said today’s top-tier AI models have improved performance by about two-thirds from November last year. The pace of new model launches also remains fast, with about 2 new models coming out every 3 days.

Actual usage patterns are different. Tunguz said 84 percent of tokens flowing through the AI model routing platform OpenRouter are handled by models other than the highest-performing ones. He said the 6 models that process most tokens deliver performance at 77 percent of the top-performing models, while costing only 2.5 percent of the price of Claude Fable 5.

As of Aug. 10, the average price of the 6 models that accounted for 80 percent of total usage over the past week was $0.5 per 1 million tokens. On the same basis, Fable 5 was $20.

Data from fintech firm Ramp also show buyers responding sensitively to price. Fable 5, launched at about $10 per 1 million tokens, accounted for 6 percent of Anthropic’s total tokens and 11 percent of revenue within a month of launch.

By contrast, OpenAI’s top paid model, GPT-5.6 Sol, accounted for about a quarter of OpenAI’s total tokens. As of July, Fable 5 was much more expensive than GPT-5.6 Sol, but its model revenue was only about 75 percent of GPT-5.6 Sol’s.

Tunguz Partner forecasts that the scale of market share shifts will shrink compared with the past each time a new highest-performing model is released.

He cited 2 reasons. One is a trend for companies to lock in contracts by concentrating spending on 1 or 2 suppliers, a pattern similar to the early cloud industry. Once a specific model establishes itself in high-value tasks, it also tends to keep that position. The other is narrowing performance gaps. Among open-source models, the best-performing model was only at 48 percent of top-model performance as of May last year, but it caught up to 80 percent by May this year.

The same trend is also appearing where applied services are built. Tunguz said many portfolio companies invested in by Theory Ventures primarily use small models, fine-tuned models and open-source models. He said they are finding an optimum based on other criteria that prioritize price over performance. He added that if users keep flocking to cheap models that are “good enough” as they are now and the shift toward highest-performing models stops, the economics of highest-performing models themselves change. He said securing market share is essential to recoup training costs that run into hundreds of millions of dollars, and that bar will rise over time.

He said there are still areas where the highest-performing models remain necessary. Fields such as software engineering design and security design, where top performance pays off, are examples. Even so, he repeatedly stressed that available data so far show the key is not the race for top performance but competition on performance relative to price.