Ask once. See how every model answers. Watch it change.
A screenshot of a chatbot shows you one answer, once. It cannot tell you whether the model would say the same thing again, whether another model would agree, or whether the answer will be different next month. Mind Polling asks for you, again and again, and keeps every reply.
1. Write the question
Ask it as multiple choice, a scale, a number or a short written answer. Add other wordings of the same question if you suspect phrasing matters (it often does). Before anything runs, you see the exact text each model will receive. For a multiple-choice question it looks like this:
Which ocean is larger?
Options:
A. The Atlantic Ocean
B. The Pacific Ocean
Answer with the letter of exactly one option, and nothing else.That is the whole prompt: no hidden instructions, no persona, no system prompt of ours.
2. Choose the models
Pick from OpenAI, Anthropic, Google and xAI, asked directly, and other labs through OpenRouter. Turn on web search for a model when the question depends on today's news, with up to 1, 2 or 3 searches per answer. Every model and its rates are on Models.
3. Set the samples and the schedule
Choose how many times each model hears each wording. Run it once, or daily or weekly at a time you pick, for up to 90 days. Start small: one run tells you whether the question works before you pay for a month of them.
4. See the cost, then fund it
The builder shows what one run and the whole poll are likely to cost before you save anything. You add funds in advance; each run sets aside what it might cost and returns the rest when it ends. See Pricing.
5. Read the results, and the answers behind them
Each model gets a chart of how its answers split, with a 95% range on every share, run after run. Open any poll's answers to read the replies, the exact prompt, the order the options were shown in and the model that actually answered. Download everything as CSV or JSON.
Three things that keep a result honest
- Options are shuffled on every call, so no answer wins by being listed first.
- Change the poll and it starts a new series, so your edit is never mistaken for the models changing their minds.
- A model that declines is shown declining. Refusals, replies that do not answer and failed calls are counted beside the answers, never hidden or turned into zeros.
What you are measuring
The models as their makers offer them to developers, under settings you can see. Not people: a share of answers is not public opinion. And not exactly what someone sees in the ChatGPT app, which adds its own instructions and memory. The methodology explains why that difference is real, with the published evidence.