𝗧𝗵𝗿𝗲𝗲 𝗠𝗼𝗱𝗲𝗹𝘀, 𝗧𝗵𝗿𝗲𝗲 𝗢𝗽𝗶𝗻𝗶𝗼𝗻𝘀, 𝗭𝗲𝗿𝗼 𝗗𝗼𝗹𝗹𝗮𝗿𝘀
I used to pay 1.50 dollars every time I asked an AI for a second opinion.
I ran three models several times a day. Each call cost about fifty cents. On my monthly bill, it looked small. In total, it was my most expensive habit. It produced no code.
Now, those same three opinions cost me zero dollars.
I use the same models and the same prompts. The difference is how I route them. I send my requests through subscriptions I already pay for. This adds nothing to my bill.
Running three models is better than running one. Here is why:
A single model is not a check. It is an echo.
If you ask one model to validate your work, it often agrees with you. It finds a way to make your logic work. You do not get a review. You get a confident version of your own mistakes.
Models have specific blind spots. These errors often cluster around a single vendor. If you use two models from the same family, they might both miss the same error.
I use three voices from at least two different lineages:
• A GPT-class model • A mid-tier Claude model • A top-tier Claude model
I ask each for a verdict and the reasoning behind it. I do not just look at the final answer. I read the reasoning.
If two models say "yes" but use different logic, I look closer. If they disagree, I have found the exact spot where I am wrong. I do not let them vote. I do not average their answers. I read where they diverge.
Why three models?
Five models is too slow. The last two calls usually just repeat what the first three said. Three models are enough to break a tie.
I stopped thinking about how many tokens I would burn. I started thinking about which subscription seat I would use.
API billing is fine for a single task. But for repetitive checks, a flat-rate subscription is better. It allows me to run checks all day without worrying about the cost.
A council validates your work. It does not invent new ideas. For a brand new problem, use one model. For daily questions like "did I miss something obvious," use three.
Three models picking at your work beats one model agreeing with you.
Source: https://dev.to/fillip_kosorukov/three-models-three-opinions-zero-dollars-41im
Optional learning community: https://t.me/GyaanSetuAi