We tracked 1M LLM API calls – 62% were using the wrong model (tokonomics.ca)

A recent analysis of one million LLM API calls revealed that a staggering 62% of developers are using advanced models like OpenAI's GPT-4o for tasks where simpler, more cost-effective alternatives would suffice. This behavior, dubbed "LLMflation," results in unnecessary costs for businesses, as many routine tasks such as classification and JSON extraction could be performed by budget models like DeepSeek V3, which is 18 times cheaper per million tokens. The average monthly AI expenditure has surged to $85,500 per company, highlighting a 36% year-on-year increase, with many organizations unaware of their cost inefficiencies. For the AI/ML community, these findings emphasize the critical need for better model selection and resource management. By implementing model routing based on task complexity and employing prompt caching techniques, companies could potentially reduce their LLM expenses by up to 95%. The analysis calls attention to the importance of benchmarking model effectiveness, particularly in a landscape where spending is growing rapidly and quality isn't continuously evaluated. Optimizing model usage not only improves cost efficiency but also ensures that teams allocate resources effectively for their AI initiatives.