LLM cost calculator

What three real jobs cost on each lab's premium and value models, computed from the live per-token rates Zeplik bills against. The token counts behind every figure are stated below, so you can scale them to your own work instead of trusting a number.

The three jobs

A short question (500 in / 500 out)

Roughly a paragraph in, a few paragraphs out. The commonest shape of turn by a wide margin.

A long document, summarised (50,000 in / 1,000 out)

About eighty pages in, a page of summary out.

A month of daily use (120,000 in / 36,000 out)

Thirty turns, each carrying a growing conversation: 4,000 tokens in and 1,200 out per turn.

Cost per model

Two models per lab: the lab's most expensive current model and its cheapest with a window of 100K or more, both derived from the registry so a new release changes this table on its own. "Raw" is the provider's own published rate; the credit figure is what Zeplik charges. 1 credit = $0.10, computed from the raw provider rate with a 1.20x margin. All 285 models carry the same figures on their own pages.

How to read it

• Input is cheap, output is not. Most models charge three to five times more for what they write than for what they read, so a long question is usually cheaper than a long answer.
• The history is the hidden cost. Every turn re-sends the conversation so far. A thread that has run for twenty messages pays for those twenty messages again on the twenty-first.
• The gap between tiers is enormous. Within one lab the premium and value rows above often differ by more than ten times, and for drafting, extraction and summarising the cheaper one usually does the job.
• Switching per task is the actual saving. Not finding the cheapest model, but using the expensive one only where it earns its price. That is one click mid-conversation on Zeplik.

Frequently asked questions

How is the cost of an AI model calculated?

Per token, in both directions. A model publishes a price per million input tokens and a different, usually higher, price per million output tokens. The cost of a job is the input tokens times the input rate plus the output tokens times the output rate, which is why a long document costs more to read than a long answer costs to write.

What is a token?

Roughly three quarters of an English word, so 1,000 tokens is about 750 words and a page of prose is around 600 tokens. Code and non-Latin scripts use more tokens per visible character.

What is a Zeplik credit worth?

1 credit = $0.10, computed from the raw provider rate with a 1.20x margin. The same balance covers chat on every model plus image, music and video generation, so there is one number to watch rather than one per vendor.

Why does the same question cost different amounts on different days?

Because the conversation grows. Every turn re-sends the history, so the input side of the cost rises as a thread gets longer. That is the single biggest reason real usage outruns an estimate made from one message.

Is a cheaper model worse?

Not for every job. The price gap between a lab's premium and value tiers is often ten to twenty times, and for drafting, extraction, classification and summarising the cheaper row frequently does the work. The expensive model earns its price on hard reasoning and long code, and the useful habit is switching per task rather than picking one.

Where do these numbers come from?

The model registry Zeplik routes and bills against, computed with the same function that charges a real turn. They are not typed into this page, so they cannot drift from what you are actually billed.