In recent months, it has become evident that while artificial intelligence expenditures are rising, companies are actively seeking more efficient ways to manage these costs. Initial enthusiasm for advanced proprietary models led to high invoices, prompting diverse optimization strategies. These include experimenting with cheaper open models running on inference platforms, smarter routing of requests to the most suitable models, and even internal knowledge sharing on cost-effective AI utilization.
Uber is a prime example of successful optimization. After exhausting its annual AI budget early in the year, engineering teams focused on cost reduction. Through a combination of strategies, including transitioning to open models for specific tasks, optimized model selection based on performance tests, using cheaper sub-agents for smaller tasks, and efficient caching, Uber managed to reduce the cost per AI request by 34% and per AI session by 52%. Despite a significant increase in overall AI usage, costs have remained stable since March. A key factor in Uber's savings was the switch to open models for certain tasks, which proved to be 2-20 times cheaper compared to the most expensive proprietary alternatives.
The social platform Pinterest reports even more dramatic savings, exceeding 90%. They leverage their own purpose-built models and carefully selected open models, which are further trained on their unique data within a secure cloud infrastructure. CEO William Ready confirmed that these adapted open models demonstrate excellent performance for their specific use cases compared to third-party closed models. Transaction costs have dropped to less than 8% of the price of comparable proprietary solutions, opening up vast opportunities for future innovation.
Telecommunications giant AT&T, with over 100,000 employees, achieved 56% savings by shifting some tasks to open models. Interestingly, the drop in AI output quality was measured at only 2%. AT&T engineers continue to utilize state-of-the-art models for complex tasks like code generation, while deploying more economically efficient open models for less demanding operations, such as summarizing existing code.
The trend indicates that while some previously popular proprietary models are noticeably becoming more expensive, open models and even some competitive proprietary alternatives are becoming more accessible. Price differences can be significant; some advanced models are up to 100 times more expensive than their more efficient open or cheaper proprietary counterparts. This compels technology companies to carefully re-evaluate their AI strategies.
Got an idea for a website or app?
Let's talk it through — non-binding, with no pressure, and a concrete next step.
A survey conducted among engineering teams at companies like Stripe, Coinbase, Uber, and Ramp confirmed that the largest savings can be achieved precisely by deploying open models. The second most effective method is intelligent model routing, followed by expense control and context optimization. These strategies are translating into real numbers, with major companies beginning to reduce their overall AI expenditures, likely not by decreasing usage, but by smart cost optimization.