Meet AI Cost: estimate what an accepted AI task really costs

An AI feature can look inexpensive on a token price table and still miss its budget. A summary might require another call, a document may cross a long-input pricing threshold, or a response may fail your acceptance rules. AI Cost helps you put those details into a comparison you can inspect.

The LLM cost calculator compares monthly API spend for Claude Haiku 5.5 and GPT-6 Luna. Its second mode uses observed samples to estimate the cost of an accepted task. It runs in your browser, requires no account or API key, and does not send prompts to either provider.

Start with the work your app actually does

Choose one task: classify a product, summarize a document, draft a support reply, or extract structured fields. Define what an accepted result means before comparing models. For extraction, valid JSON is only one requirement; the values must also match the source.

Collect representative usage for both models. The same text can produce different token counts, and different reasoning settings can change billed output. Use the provider’s billed usage rather than counting visible words.

Open the calculator and replace the demo request volume, days per month, input, and output. The defaults are illustrative. They do not represent an industry average or a measured quality score.

Walk through a small monthly estimate

Suppose you make 1,000 requests a day for 30 days. Each request uses 2,000 uncached input tokens and 500 billed output tokens. With short-input Standard prices of $0.10 per million input tokens and $0.50 per million output tokens, the arithmetic is:

Input per call:   2,000 × $0.10 / 1,000,000 = $0.00020
Output per call:    500 × $0.50 / 1,000,000 = $0.00025
Monthly total:  30,000 × $0.00045           = $13.50

Both models give the same estimate for these deliberately equal inputs. That tells you the base price is not the deciding factor for this scenario. It says nothing about which model gives a better answer.

See the Haiku 5.5 vs Luna 6 comparison for the dated price table and longer-request examples. The official Haiku model page and GPT-6 Luna model page are the primary sources for the listed rates.

Keep long requests separate

Averages can conceal an expensive part of your workload. If most calls contain a short label but some contain a large document, create a separate request class for each. AI Cost chooses a pricing tier for each class before adding the costs.

In the current snapshot, Haiku changes tiers above 100,000 input tokens; Luna changes above 272,000. The higher rates apply to the entire request, including output. Exactly at the threshold still uses the shorter-input tier.

Cache reads and writes count toward full input length. Reusing a prompt can reduce the applicable token charge without placing a long request into a shorter tier.

Count cache writes as well as reads

Open Cache & extra costs when you have cached usage. Enter three separate input categories: uncached input, cache reads, and cache writes. A token belongs to one category, so do not enter total input as uncached and then add cached input on top.

For Haiku, select the cache-write duration that matches your workload. For known per-call fees, enter an extra cost explicitly. If you use a contract or third-party rate, enable custom pricing; the calculator marks the result as user-defined and replaces the official tiers with your flat rates.

The methodology lists the charges that are not automatically included. Tools, tax, regional processing, other service modes, and account credits can make an invoice differ from an estimate.

Measure accepted tasks without counting retries twice

Imagine you start 1,000 tasks, make 1,300 billed calls including retries, and finally accept 900 tasks. If the whole sample costs $4.50, the cost per accepted task is:

$4.50 / 900 accepted tasks = $0.005 per accepted task

The observed calls per task are 1.3, but you do not multiply $4.50 by 1.3 again. The sample total already includes those calls. Each task counts at most once as accepted, even if three separate calls were needed to finish it.

At 30,000 started tasks per month, the same sample projects $135 of included spend and 27,000 accepted tasks. These are mathematical projections under an unchanged task mix, not predictions of a model’s future behavior.

Enter the same kind of sample for the other model. Set your minimum acceptance rate and, if relevant, a maximum average latency. A missing latency measurement cannot establish that a model meets a latency limit. If no task is accepted, the calculator shows that no finite accepted-task cost is available.

Save an assumption, not a secret

Use Save on this device to store one plan in the current browser. Use Copy share link to share its numeric settings. A link does not include prompts, API keys, or logs, but anyone with it can read the included workload and cost assumptions. Share only numbers you are comfortable disclosing.

Saved plans are recalculated with the current deployed price snapshot when you load them. AI Cost displays a notice if the saved version differs. The site also shows an age warning when listed prices are more than 30 days old.

Start with the calculator, check the model comparison, and keep the calculation method beside your estimate. If a price or result looks wrong, email hi@sevu.net with the official source and non-sensitive values needed to reproduce it.