AI industry explores new pricing models to manage token and task costs

The artificial intelligence industry is exploring new methods to manage and bill for the high costs of model consumption. While traditional pricing is based on token usage, new approaches are emerging to address the efficiency gap between different model tiers.

Snowflake reported that using a frontier model for a million-row data task cost approximately $52,000, whereas a smaller, open-weight model performed the same task for $900. However, Databricks found that cheaper models can sometimes be more expensive in practice. Their testing showed that less capable models may require more attempts to complete a task, leading to higher overall costs per task due to wasted tokens. In one instance, Anthropic’s Sonnet 5 model cost more per task than the more expensive Opus 4 due to lower output quality.

In response to these cost complexities, OpenAI is experimenting with outcome-based billing for enterprise customers. Instead of charging for every token consumed, this model aims to bill customers only when an AI successfully completes a task. This shift presents technical challenges, as determining the exact moment of success for complex AI agents—such as those performing coding or browser-based tasks—is significantly more difficult than simply counting tokens.