Pricing model

Every request through the gateway is priced from one thing: the token usage that the provider reports back. There is no per-request fee, no minimum charge, and no flat rate per model. Cost is computed from your actual input and output - nothing else.

What a token is

A token is roughly 4 characters of English text, or about three-quarters of a word. "Hello, world!" is around 4 tokens. A typical paragraph is 80-150 tokens. Code, JSON, and non-Latin scripts can be denser or sparser depending on the tokenizer the model uses.

Every text model in the world prices on tokens - this is not specific to Aivene.

Input vs output

Each model has two prices, both quoted per million tokens (MTok):

This is why a request that returns a long essay costs much more than one that returns a single sentence, even when the prompt is identical.

The formula

The base cost of a single request is:

cost = (input_tokens  / 1,000,000) * input_price
     + (output_tokens / 1,000,000) * output_price

That is the entire model. Everything else - caching discounts, tool fees, discount tiers - is a modifier on top of that formula.

Worked example

Say a model is priced at $3.00 / MTok input and $15.00 / MTok output. You send a request with 2,000 input tokens, and the model replies with 500 output tokens:

input cost  = 2,000 / 1,000,000 * $3.00  = $0.006
output cost = 500   / 1,000,000 * $15.00 = $0.0075
total       = $0.0135

One request, just over a cent.

If your prompts and responses stay roughly that size, a $10 balance gives you around 740 requests on that model. Swap to a model priced at $0.20 / $0.80 per MTok and the same $10 stretches to roughly 14,000 requests - same dollars, different model, ~19x the volume.

This is why we cannot publish a "requests per dollar" number. It is a property of your prompts, not the gateway.

Prompt caching: cheaper repeat input

When you send the same prefix to a model more than once, most providers cache it and bill those repeat input tokens at a steep discount. Cached input is reported in the response as:

{
  "usage": {
    "prompt_tokens": 2000,
    "completion_tokens": 500,
    "prompt_tokens_details": {
      "cached_tokens": 1800
    }
  }
}

Of the 2,000 input tokens above, 1,800 hit the cache and are charged at the discounted rate; only 200 are billed at the full input price. See Prompt Caching for which providers support this and how big the discount is.

Built-in tool fees

If your request invokes one of our built-in tools - web search, web fetch, or code interpreter - that tool has its own per-call price that is added on top of the model tokens for that request.

total_cost = model_cost + tool_cost

Tool prices are listed on each tool's page under Built-in Tools. You only pay the tool fee when the model actually calls it; defining a tool but never invoking it costs nothing extra.

Other modalities

Tokens are not the only billable unit. Depending on the endpoint:

Whatever the unit, the cost still comes from numbers the provider returns in the response, multiplied by the model's published per-unit price.

Reading what you paid

Every response includes a usage object describing exactly what was counted. In the Console:

• Logs shows the per-request token breakdown and the dollar amount the gateway subtracted from your balance.
• Usage rolls those numbers up by model, key, and date so you can spot which model or which key is eating the budget.