How these AI task costs are calculated

Exchange rate

The RBA table for 19 Oct 2026 lists the United States dollar at 0.6975 per Australian dollar. AUD figures on this site equal the USD figure divided by 0.6975 (multiplied by 10000/6975) and are rounded half up. The RBA page says the series should not be relied on for regulatory or commercial purposes. Vendors below do not publish these AUD amounts. Reserve Bank of Australia, read 2026-10-20. Rate date on the table: 2026-10-19. Quote used: 0.6975 United States dollars per Australian dollar.

By hand

Clerks—Private Sector Award 2020 [MA000002]. Adult Level 2, Year 1, minimum hourly rate for ordinary hours (clause 16.1, Table 3). A$29.45 per hour, in the award from 2026-07-01. The award page incorporates amendments up to and including 1 July 2026 (PR799283, operative from the first full pay period on or after 1 July 2026). Table 3 lists Level 2, Year 1 at $29.45 an hour and $1,119.10 a week. This is the minimum hourly rate for ordinary hours. It is not the casual rate ($36.81), not overtime, and not a loaded cost with superannuation, leave, or payroll tax. Fair Work Ombudsman, Clerks—Private Sector Award 2020, read 2026-10-20.

Minutes per task are an assumption, not a time study. Each task page labels them "assumption, edit it". The hourly wage starts at the award rate and can be edited too. A USD amount in that calculator converts the edited AUD wage with the RBA rate above.

Text size

OpenAI's token help article says that for English, 100 tokens are approximately 75 words. OpenAI help: tokens, read 2026-10-20. Word counts on this site are chosen so that conversion lands on a whole number of tokens (words × 100 / 75). The same count is applied to Anthropic and Google. Their tokenizers can differ. Reasoning tokens are not added, because a count for these scenarios is not published.

Sort order

Inside a task, rows are sorted by the calculated USD amount, lowest first. Rows with the same calculated price share a rank, shown as =1 rather than 1 then 2. That is not a quality ranking. A sponsored label would not move a row or change a figure.

On a task card, Cheapest is that first row. Typical is the median listed row when its price is higher. It is one of the listed rows, not an average.

Left out

• Monthly seat plans, including ChatGPT and Claude subscriptions, are not split into a per-task price. A per-task figure would require a task count the vendor does not publish.
• OpenAI image models are omitted. The image guide prices output in tokens and says the token count depends on quality and size. The pricing page does not fix a token count for GPT Image 2.5, so this sheet does not invent one.
• Transcription rows whose per-minute figure is in a column the OpenAI pricing page heads “Estimated cost” are omitted. GPT-Transcribe is included because its model page states a price per minute.
• Free tiers are not turned into a per-task price. The pages read here do not say how many of these scenarios a free allowance covers.

What this is not

Not a quote from a vendor. Not a price in Australian dollars on the vendor’s invoice. Not health, legal, medical, tax, or financial advice. AI Cost Per Task does not say a setup is safe or compliant.