I want to pay for a completed task, and a token discount alone cannot tell me what that costs. A cheaper call can still fund an endless meeting between agents. The calculation needs to reach a deliverable somebody can accept.
On August 21, 2026, OpenAI cut GPT-5.6 Sol's price to $4 per million input tokens and $20 per million output tokens, as promotional pricing. A discount like that makes it tempting to split work across parallel agents, each responsible for a separate part.
The trouble starts with how the work gets divided. If each agent must discover the same information before proceeding, part of the budget buys repetition. Then somebody has to compare the answers and resolve disagreements. That can justify its cost in a difficult investigation. For a simple assignment, coordination can end up outweighing useful work. Parallel execution needs a reason beyond an available slot.
The same applies to choosing the model. An economical choice starts with the result required and the cost of repairing a poor attempt. A cheap model that leaves work unfinished may need another run with additional context.
I prefer splitting the work into activities that return something useful on their own. A search can bring back evidence while another part advances. Two agents renegotiating the same decision at every step need a different arrangement. The number of executors becomes interesting only after each responsibility is clear enough to explain.
Since the price is promotional, a configuration chosen because of it needs another look when the promotion ends. Multiplying the new price by imagined consumption helps planning; savings show up when I observe the execution. The rule I use is cost per accepted deliverable, counting discarded attempts and the attention spent coordinating.