List prices in US dollars per million tokens, with country of origin, licence, and what each vendor actually says about Australian availability and data residency. Click any row for the detail and the source.
Every row verified 7 September 2026 against the vendor's own pricing page. No Australian-headquartered frontier model provider was found with published API pricing.
| Model | Origin | Licence | Context | Input | Output | Cached in | Batch in |
|---|
Prices per million tokens, standard list tier, no enterprise discount. Context in thousands of tokens. A dash means the vendor publishes no rate for that line. It is not zero and it is not an estimate.
Nine of the twenty-odd providers here are Chinese, and most of those publish open weights under a licence that permits self-hosting. Every one of their hosted APIs runs from China or Singapore. Of the American providers, only Amazon and NVIDIA-on-Bedrock name an Australian region and rate. Google runs a Sydney Vertex region but does not state per-model availability on its pricing page. Mistral's Apache 2.0 models are the cheapest weights a business can legally run on its own Australian hardware.
For an Australian business with a data-residency obligation, the practical shortlist is much shorter than the price table suggests: Bedrock in Sydney, a Vertex deployment you have confirmed yourself, or open weights on your own infrastructure. Everything else means your prompts leave the country, which is a decision to make deliberately and record, not one to discover during an audit.
