Knowledgebase

Each model has three prices, in credits per million tokens: fresh input, cached input and output. A request costs:

(fresh input × input price + cached input × cached price + output × output price) ÷ 1,000,000

rounded up once per request. Each answer reports its cost in usage.credits, and the usage table on your Virsilipas AI key page adds them up by day and model. The prices are on the wiki: Models and what a request costs.

Why did that cost so much?

• Output costs more than input, and reasoning counts as output. A -think model can write a long reasoning text before the answer.
• Coding assistants send a lot. Claude Code sends its full instructions and tool descriptions with every request; in our test, a task with a single shell command cost about 5,000 credits on default. A long agent session reads many files and costs much more.
• A conversation grows. Every new message resends the whole conversation so far.

What is not charged

A refused request costs nothing, and only an answer that is delivered is charged. A request can never spend more than your balance.

Why do I need more balance than a request costs?

While a request runs, the service holds enough credits for the longest answer it could give, counting the prompt generously. The unused part returns when the request finishes. When your balance is low, the longest answer allowed shrinks to what your balance covers, so answers can stop early.

Spending fewer credits

• Use default or an -instant alias unless the task needs more.
• Start a new conversation instead of continuing a very long one.
• Set max_tokens to cap the length of an answer.
• For tools that resend the same text, try alt, which is very cheap on repeated (cached) input.