1300 166 138 Tell us what you need
News / OpenAI

OpenAI’s API pricing explained for Australian businesses

By Monika Tantau • Published 3 weeks ago
A developer with lines of code reflected in his glasses works at a screen, the kind of team that pays for OpenAI's API by the token

Key takeaways

• GPT 6.1 Sol costs about $2.85 per million input tokens and $14.25 per million output tokens
• A new 12 hour prompt cache, live since 20 October, costs 2.5 times the input price to write
• Data residency adds 10%, and Australia gets onshore storage but not onshore processing
• Existing fine tuning customers can no longer start new jobs from 6 January 2027

Most Read

OpenAI charges for its API by the token, in US dollars, and the bill depends on more than the model. Speed tier, caching and where data is kept all change the price. This explainer converts the current OpenAI API pricing into Australian dollars at the Reserve Bank’s rate on 22 October.

The only genuinely new item is a 12 hour prompt caching option, live since 20 October. Fast mode, Ultrafast, the 10% data residency uplift and the end of fine tuning for new users all date from earlier in the year, but they matter just as much to an Australian budget.

What each GPT 6 model costs

The prices below are per million tokens on the Standard tier. Because OpenAI bills in US dollars, what an Australian team pays moves with the exchange rate.

GPT 6 Astra, the most capable model, costs about $14.25 per million input tokens and $71.25 per million output tokens.

GPT 6.1 Sol, aimed at coding and professional work, costs about $2.85 per million input tokens and $14.25 per million output tokens.

GPT 6 Luna, the cheapest of the three, costs about 14 cents per million input tokens and 71 cents per million output tokens.

Batch and Flex processing halve those prices, and Fast mode doubles them. Ultrafast is listed only for GPT 6 Astra and GPT 6.1 Sol, at about $85 and $17 per million input tokens. Our comparison of GPT 6 Astra, Sol and Luna covers which model suits which job.

Hands hold a tablet showing a workflow dashboard in a server room, a reminder that how long a prompt stays cached changes what each request costs

The new 12 hour cache

Prompt caching lets a business pay once to store the repeated start of a prompt, such as instructions or a policy manual, and then pay far less each time it is reused. Until now the cache lasted 30 minutes by default. Since 20 October developers can choose 12 hours instead.

A 30 minute cache write costs 1.25 times the normal input price and a 12 hour write costs 2.5 times. Reading from either cache costs the same, which on GPT 6.1 Sol is 5% of the normal input rate. On that model, a 12 hour write costs about $7.12 per million tokens.

Take a support assistant on GPT 6.1 Sol with 50,000 tokens of policies at the start of every prompt, used about once an hour over a ten hour day. With 30 minute caching the context expires and is written ten times, for about $1.78. Uncached it costs about $1.42. One 12 hour write plus nine reads costs about 42 cents. For an AI chatbot answering from long documents all day, that gap adds up.

What it means for Australian data

OpenAI charges a 10% uplift on its data residency endpoints for models released on or after 5 March 2026, which covers the whole GPT 6 family. That lifts GPT 6 Astra input from about $14.25 to $15.67 per million tokens. Access to data residency also goes through OpenAI’s sales team.

For Australia the detail matters. OpenAI’s own data guide lists Australia for regional storage but not regional processing, so stored data can stay onshore while prompts are still processed offshore. Our guide to AI data residency in Australia explains what that means under the Privacy Act and the US CLOUD Act.

The fine tuning deadline

OpenAI stopped new organisations from fine tuning in May. Its deprecations page sets 6 January 2027 as the date existing customers can no longer create new training jobs. Fine tuned models keep working until their base model is deprecated, so nothing breaks in January, but nothing new can be trained either.

Teams that rely on fine tuning have until then to test whether better prompting, retrieval or another provider can replace it. Starting now leaves time to compare quality and cost.

How to keep the bill down

Since 21 October, admins can choose which service tiers each project may use. That is a simple way to stop a test project switching on Ultrafast, which on GPT 6 Astra costs six times the Standard rate.

Matching cache length to traffic, locking down expensive tiers by project and budgeting 10% for residency will usually save more than switching models. The token price is only the starting point of an OpenAI bill.

Scroll to Top