One Smart AI Could Manage Several Cheap Ones
Atallah says developers should try having cheap, open models do repetitive chores while the costly frontier model handles judgment calls.

20VC with Harry Stebbings · OpenRouter CEO: Why Chinese Open Models Are Beating the US | Why Enterprises Fear OpenAI & AnthropicAn idea worth testing for getting work out of artificial intelligence is not to pick one best model. It is to hire one best model as the boss and staff it with cheaper ones, according to Alex Atallah, co-founder and chief executive of OpenRouter, a marketplace where developers buy access to many AI models.
When asked about one powerful frontier model, the most capable and expensive kind, directing four open models, Atallah said that structure is a great architecture for developers to explore.
In his version, the frontier model acts as the orchestrator. It farms out routine jobs to several very low cost, open models, whose trained settings are public so others can run them.
The division of labor turns on how predictable the job is. Open models are especially good at deterministic tasks, work where the shape of the answer is known in advance and the problem has been solved before. Classifying text is one example.
The architecture uses what Atallah calls a sub-agent server tool, software tuned to help models call on other models. When a smaller chore comes up, the orchestrator passes it to a cheap sub-agent, reads the result when it comes back, and keeps working on the larger, open-ended problem.
Atallah's advice for builders is to use the top model for open-ended work where the answer is unclear. In his framing, inexpensive models handle the repetitive steps around it, then the orchestrator reads the results and continues the unknown work that requires judgment.
Questions people ask
What is a frontier model?
It is the term for the most capable, cutting edge models from leading labs, such as OpenAI's GPT series or Anthropic's Claude. They tend to be the smartest on open-ended reasoning, and also the most expensive to run.
What does open-weight or open model mean?
It means the model's trained settings, called weights, are publicly released so others can download, study, and run it. That openness creates competition among hosting providers, which usually drives the price per task down sharply.
Why is classification so much cheaper to outsource?
Classification has a fixed answer format, such as spam or not spam, urgent or routine. Because the pattern is narrow and repeats, a smaller model can handle it with far less computing power than an open-ended reasoning task requires.



