LLM Cost Estimator
Estimate multi-model API cost from tokens or pasted prompts.
OpenScore prompt drafts side by side with weights you control, before any model is called.
Compare two or three candidates side by side.
Score clarity, constraints, few-shot, and output format.
Heuristic checks run privately with no model call.
Share results or save the last rubric in localStorage.
Three people on the team each rewrote the support prompt, and nobody wants to burn an afternoon of eval credits comparing them blind. A quick structural check shows which draft is missing an output contract or failure behavior, and that settles the first round of the debate in seconds.
Paste variants into boxes A and B, and optionally C. Pick a preset (general purpose, code generation, SQL generation or customer support) and adjust the six criteria: clarity and specificity, constraints and guardrails, few-shot examples, output format contract, role and context, and domain fit. Each criterion has its own weight and an on or off toggle. The tool scores every criterion from 0 to 100 and reports a weighted total, with notes explaining what earned or lost points, such as a missing numeric limit or no labeled Input and Output pair.
The scoring is heuristic. It looks for patterns like 'You are', 'never', word limits, JSON keys and fenced examples, so it rewards prompt structure rather than predicting model quality. A tip: if two variants tie, read the notes rather than the totals, because the gaps they point to are what you would fix next. When a draft needs a structural rewrite, the Prompt Optimizer adds missing sections, and Prompt Generator starts a clean version from role, task and format.
Scores are calculated in your browser. Save rubric stores your weights in localStorage so they are there next visit, and the scorecard copies or downloads as plain text.
Keyboard shortcut
Answers for this tool. For site-wide help, open the FAQ hub.
No. It checks how a prompt is written, not what a model produces from it. Use it to catch structural gaps cheaply, then run your eval set on the variants that survive.
Each enabled criterion is scored from 0 to 100, multiplied by its weight, and the sum is divided by the total weight of enabled criteria. Unchecking a criterion or setting its weight to zero removes it from the total.
The heuristic looks for labeled Input and Output lines, fenced blocks and words like example. A worked example written as plain prose scores lower because there is less structure for the model to imitate.