The right model.
For every dollar.

MPortal decides fresh every time, from data that moves underneath it. Three things change the answer:

• Prices move. We pull real per-token prices from OpenRouter and refresh them through the day. Models cost what they cost that hour — when one provider drops its price or another adds a model, the cheapest option that still clears the quality bar changes with it.
• The prompt is read, not remembered. Your text is scored for task type and difficulty each time. Reword it, add a file, or ask it as a follow-up inside a project — where earlier turns are replayed as context — and it can score harder, which lifts the quality bar and lands you on a stronger model.
• Models come and go. Anything unavailable at the aggregator is skipped for that request.

Same prompt, same mode, same prices, same catalogue → same model. Open Receipt › Models considered on any answer and you'll see the exact table it chose from, with the price and quality score of every candidate.

It isn't a per-request fee. It's measured usage:

cost = (input tokens × input rate) + (output tokens × output rate)

Every model publishes its own two rates, quoted per million tokens. A token is roughly ¾ of a word. We then apply your plan's markup — Basic ×1.25, Plus ×1.15, Enterprise ×1.10; Solo runs at cost — and that number is what lands in your weekly budget.

Because the token counts are different each time. The receipt is built from what the request actually used, not from an average, and two things vary:

• The model's answer length. Language models don't produce identical output twice. Ask the same question and one run writes four paragraphs, the next writes seven. Output tokens are usually the larger half of the bill, so the total moves with it.
• What went in. Your prompt length, any attached files, and — inside a project — the replayed conversation turns and the project's instructions all count as input tokens.

So two DeepSeek runs at the same rate can still differ: one wrote 400 tokens, the other 700. The per-token price didn't change; the amount of work did.

No — it's a ceiling, not a price. It tells the router "don't pick anything that could cost more than this". Almost every request comes in at a small fraction of it. The receipt shows the estimate made before sending and the real figure after.

A little, and never invisibly. Continuing a conversation sends the recent turns back to the model so it knows what you're talking about, and those turns are input tokens like any others. MPortal prices the whole context — instructions, replayed turns, your new message — before it routes, so memory is inside the estimate and inside your budget. Start a new chat any time to drop it.

It shouldn't, and the receipt will tell you why it did. Difficulty is judged from the prompt: a short question that happens to be hard can read as easy. Two fixes — say what you want ("production-grade", "step by step", "check the edge cases"), or switch the mode to Best for that request.