What an AI task costs on the DipoleML API, worked to the cent

Per-token prices are easy to publish and hard to picture. This piece prices three everyday tasks on the DipoleML API, on both Raven models, from the prices on our pricing page. The token counts are our assumptions for each task, stated as we go; the arithmetic is the API's own. The requests are Chat Completions in the OpenAI SDK format.

How is a request billed?

Calls with an API key are paid from prepaid credits, per token, at the prices of the model you call:

Each request costs its new input at the input price, its cached input at the cached price and its output at the output price:

cost = (input - cached) x input price
     + cached x cached price
     + output x output price

(prices per 1M tokens)

Two details shape the sums below. Reasoning tokens are output tokens: they count toward max_tokens and are billed at the output price. And cached input is the part of your input the model reports as already cached, in usage.prompt_tokens_details.cached_tokens. It is applied on its own, with no flag to set, at up to 90% less than the input price.

What does a code review cost?

Say you send a pull request's diff with the files around it, 30,000 tokens in all, and ask for a review. The model reasons for 2,500 tokens and writes a 1,500-token review, so the output is 4,000 tokens. Nothing is cached, because the code is new each time. Here is what a thousand such reviews cost:

One review costs $0.0250 on Raven Flash and $0.0625 on Raven Max.

What does a support reply cost?

A support assistant sends the same long instructions with every ticket: say 10,000 tokens of policies and product facts at the start of the request, then the customer's thread, 1,000 tokens. The reply is 600 output tokens, reasoning included. After the first reply, the model reports the repeated instructions as cached, so they are billed at the cached price:

One reply costs $0.0025 on Raven Flash and $0.0063 on Raven Max. For 10,000 replies, that is $25.00 and $63.00. Without the cache, the same replies would cost $70.00 and $175.00.

What does it cost to read a long document twice?

Say you send a 150,000-token contract with a 400-token question, then send the same contract again with a second question. Each answer is 2,000 output tokens, reasoning included. Both requests fit in the 1M-token context window of either model. On the second request, the model reports the contract as cached:

Both questions together cost $0.0929 on Raven Flash and $0.2330 on Raven Max. Without the cache, the second question would cost as much as the first, and the two together $0.1604 on Raven Flash and $0.4010 on Raven Max.

Where do I read the token counts?

In the usage of every answer: prompt_tokens for the input, completion_tokens for the output with its reasoning, and prompt_tokens_details.cached_tokens for the cached part. The same formula in Python, with the OpenAI SDK pointed at our base URL:

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.dipoleml.com/v1",
    api_key=os.environ["RAVEN_API_KEY"],
)

# US dollars per 1M tokens: input, cached input, output
PRICES = {
    "Raven Flash": (0.50, 0.05, 2.50),
    "Raven Max": (1.25, 0.13, 6.25),
}

MODEL = "Raven Flash"
resp = client.chat.completions.create(
    model=MODEL,
    messages=[{"role": "user", "content": "Review this diff: ..."}],
)

u = resp.usage
details = u.prompt_tokens_details
cached = (details.cached_tokens or 0) if details else 0
inp, cac, out = PRICES[MODEL]
cost = ((u.prompt_tokens - cached) * inp + cached * cac + u.completion_tokens * out) / 1_000_000
print(f"${cost:.4f}")

How do I keep the cost down?

• Send the work that fits it to Raven Flash. Raven Max's input and output prices are 2.5 times Raven Flash's.
• Put what repeats at the start. Instructions and documents that open every request are where the cache applies.
• Let simple jobs think less. With reasoning_effort set to low, the model spends fewer output tokens on reasoning.
• Give each app its own key. A spending cap on each key keeps one app from using the whole balance; see Management Keys.

How do I pay for it?

With prepaid credits, in packs of $5, $10, $20 and $50. Credits never expire, every API key on your account draws on the same balance, and they are separate from every plan: you do not need a plan to call the API, and a plan never uses your credits. When the balance runs out, requests are refused with 402 until you add credits. The details are in the API Credits docs, and the models and formats on the API page.

Product names are trademarks of their owners and are used here only to identify the products.