AI MODEL
ROUTER
Paste a prompt. See the cheapest model that can still do the job.
Frequently asked questions
How does the AI model router decide?
It scores the prompt on measurable features: the kind of task, whether it asks for step-by-step reasoning, strict rules, listed requirements, code, length and language. Those add up to a difficulty score from 0 to 100, which sets the minimum benchmark score a model needs in the matching category. The cheapest model reaching that level is recommended — preferably one that reaches it with room to spare rather than only just. Every signal and the points it added are shown with the result.
Are the cost savings real?
The cost side is exact: your prompt's token count is measured with the GPT-4 tokenizer, adjusted for each model's own tokenizer, then multiplied by that model's published price. The only assumption is the output length, which you set in the form yourself. Which model will suffice is an estimate — that is why every result carries a confidence level.
Do you send my prompt anywhere?
No. Tokenisation and routing run entirely in your browser. The prompt is never uploaded, stored, or logged, and no account is required.
Why does it sometimes suggest a more expensive model than the cheapest option?
Because a model that only just meets the requirement is the riskiest pick: if the difficulty estimate is even slightly off, you are left with one that cannot do the job. So we prefer a model with room to spare. When that room costs far more — sometimes thirty times more — we recommend the cheaper model instead and say plainly that it is borderline. Cheaper borderline options always appear separately under “cheaper, but riskier”.