Fireworks AI Co-founder Chenyu Zhao Proposes a Two-Model World: Companies Spending Over $10,000 Monthly on Tokens Should Consider Switching to Open-Source Models
Fireworks AI co-founder Chenyu Zhao outlined a specific framework for enterprise AI model selection at the Slush conference in Helsinki: when monthly token spending exceeds $10,000 to $100,000, companies should consider switching to open-source models. The future landscape involves frontier models handling high-level planning and open-source models executing tasks, with routing distributing work based on task cost rather than per-token price.
Chenyu Zhao, co-founder of Fireworks AI, presented a concrete financial threshold for enterprise AI expenditure at the Slush startup conference in Helsinki, Finland: when user scale exceeds 10,000 to 100,000 and monthly token spending reaches $10,000 to $100,000, it is time to seriously consider switching from frontier models to open-source models. His core argument is not that open-source models have caught up to frontier models in capability, but that most enterprise workloads do not require the general intelligence level provided by frontier models.
Zhao described the future landscape as a 'two-model world': frontier models are used for high-level planning and complex reasoning, while open-source models handle daily execution, operating in parallel. The key mechanism connecting them is routing—assigning tasks to the most suitable model based on cost rather than per-token unit price. He stated that the industry is no longer focused on the cost per token, but on the total cost to complete a single task. This distinction has practical implications for enterprises: an open-source model may have a lower unit price, but if it requires more tokens to complete a task, the total cost might not be lower; conversely, a frontier model with a higher unit price might be cheaper overall if it completes the task in one go.
Zhao’s premise is that the capability gap between open-source and frontier models is currently at most six months, and likely closer to three. He noted that a wave of excellent open-source models emerged following DeepSeek, significantly narrowing the gap from what was once one to one-and-a-half years. However, this assessment comes from his own evaluation and is not independently measured. Notably, Anthropic released Opus 5.5 in September, which reduced costs for typical workloads by 40%, lowering input and output token prices to $4 and $20 per million, respectively, and cache reads to $0.20 per million. As frontier models themselves continue to drop in price, Zhao’s suggested switching threshold may shift upward over time.