Need to avoid errors? Don't hire an expert.

On Who Wants to Be a Millionaire, "Phone a Friend" got it right 65% of the time.
"Ask the Audience" hit 91%.
The expert lost to the crowd.

Cognizant just proved this works for AI reasoning too.
They solved a task requiring 1,048,575 correct decisions.
With zero errors.
State-of-the-art reasoning models failed after ~200 steps.
But they did it with a bargain model.

gpt-4.1-mini at $1.60/M tokens
beat o3-mini at $4.40/M tokens.
Total cost for the million-step task: $3,500 vs. $9,400.

How’d they do it?

1. Molecular decisions
Break every task into the smallest possible unit.
One agent, one micro-decision.
No agent sees the whole puzzle.

2. Massive redundancy + voting
Multiple agents solve each micro-step independently.
First answer to win by k votes takes it.
Ask the Audience, at every single step.

3. Red-flagging confusion
If an agent's output is too long or poorly formatted, toss it.
Bad formatting correlates with bad reasoning.
Don't try to salvage the answer.

Today, when you give agents bigger chunks of work, the cost scales exponentially.
At one step per agent, cost scales log-linearly.
The difference at a million steps? Orders of magnitude.

Stop asking "what's the smartest model for this process?"
Start asking "what's the smallest reliable unit of work, and how do we vote on it?"

Cheap models.
Tiny tasks.
Vote on everything.

You don't need GPT-5 to run your supply chain.
You need a system that optimizes for cost-per-step divided by accuracy-per-step.

Need to avoid errors?
Don't hire an expert.
Hire a crowd.

—

Source: Solving A Million-Step Llm Task With Zero Errors, Meyerson et al, Nov 2025