How cost estimates work, and why some replies show none

Under each reply in Max Harness there is a small line such as 1,240 in · 380 out · est. $0.0031. Sometimes the last part is missing. That is deliberate.

Two inputs: tokens and price

A cost estimate needs two numbers.

Token counts come from the provider. At the end of a response, the API reports how many tokens it read and how many it wrote. Max Harness displays those figures as they arrive. It does not try to count tokens itself, because every model family counts differently and a guess would be wrong in ways that are hard to notice.

Price per token is the harder part. There is no standard way for a provider to tell software what a model costs. So Max Harness uses a price only from sources it can name:

• OpenRouter includes per-token prices in its model list and reports the charged amount for each request. When that is available, the figure shown is the reported cost, without the "est." prefix.
• Anthropic prices are taken from Anthropic's published list. The date of that list appears in the tooltip.
• You can enter a price for any model on the Providers page.

For every other model, there is no dollar figure. You still see the tokens.

Why not ship a big price table?

It would be out of date quickly, and stale prices are worse than no prices. Model prices change, new models appear weekly and many providers charge different rates by region or tier. A table that is right for most models most of the time produces confident numbers that are sometimes wrong, and nothing on screen tells you which ones.

Leaving the figure out is less convenient and more honest. If you care about the cost of a particular model, entering its price takes a few seconds, and from then on the estimate is based on a number you chose.

Where estimates and invoices differ

Even with a correct list price, the estimate is the token count multiplied by that price. Real bills have more moving parts:

• Cached input. Several providers charge much less for input they have seen recently. The estimate uses the full input price, so it tends to run high on long conversations.
• Reasoning tokens. Reasoning models bill for thinking that you do not see. Providers include these in the output count, and Max Harness adds them where they are reported separately.
• Long-context pricing. Some models charge a higher rate once a request passes a size threshold.
• Tool use. Each tool step is another request that resends the conversation. The usage line under each step covers that step only.
• Provider-side tools and search. Models that search the web themselves may add per-search fees that do not appear as tokens.
• Discounts. Batch pricing, committed-use plans and negotiated rates are invisible to the software.

For all of these reasons the figure is labelled as an estimate. Treat it as a way to compare models and spot expensive habits. Use the provider's billing page for the amount you will actually pay.

Practical uses

• Compare two models on the same prompt with the comparison view and look at both the answers and the usage lines.
• Watch the input count climb through a long conversation. When it is large, starting a fresh chat with a short summary is cheaper.
• Check what an attachment costs. A long PDF sent as text can be tens of thousands of input tokens on every message that follows.