The next AI race may be won by model teams, not a single champion

Ethan Mollick was responding to the claim that OpenAI and Anthropic compete at the cheaper end of the market as well as the frontier. He argued that the two vendors may be best placed to optimize cost against capability by combining agents of different ability levels that work together inside their own systems. This was an argument and a forecast, not a product announcement.

Value is moving from the model to the lineup

Mollick's point changes the unit a customer buys. Instead of one best model, a task might use a cheaper worker, a stronger planner and a separate evaluator. The provider then sells more than the intelligence of one run. It sells the method for dividing work and selecting the result.

That logic gives OpenAI and Anthropic an advantage through their model families, training data and runtime. The post provides no measurement, price or description of a specific harness, however. Their presumed lead remains a hypothesis.

The right role can matter more than the strongest model everywhere

The 2026 AgentCARD preprint tests teams with different models and roles. Its authors report up to 44% higher accuracy than cost-equivalent homogeneous teams, and in some hybrid deployments performance matching the strongest homogeneous team at up to 12 times lower cost per task. They also find that the best assignment depends on the domain. Some domains are planner-bottlenecked, while others depend on the executor.

A separate preprint on selecting among agent outputs found an advantage for a mixed team containing a weaker model in one experimental setup. The authors explicitly label that result exploratory. The shared lesson is more restrained than the swarm narrative: a good selector and correct roles may be worth more than another expensive agent.

More agents can multiply both the bill and the errors

An agent team does not automatically create team intelligence. Homogeneous models can repeat the same mistake, while synthesizing several answers may add latency without improving the result. Every extra role also consumes tokens and makes responsibility harder to trace.

A proprietary harness adds commercial risk. If quality lives inside a vendor's private routing, evals and coordination, customers struggle to compare individual models or replace the platform. A clever lineup can become as strong a lock as the frontier model itself.

Whole-task evals will decide, not member leaderboards

The useful test must compare complete workflows on real tasks: cost per accepted result, latency, human interventions and errors after completion. A single-model benchmark cannot show whether a cheap executor paired with a strong evaluator beats one expensive run.

OpenAI and Anthropic will have a real advantage if they publish repeatable results across roles and let customers alter the team composition. Otherwise cost-to-intelligence optimization remains a black box whose clearest feature is the bill issued by its maker.

Lilith's verdict

The future winner may not have the smartest agent. It may have the best casting, direction and someone who cuts a bad performance before it reaches the customer.