Australian inference.
The details matter.

gpt-4.1-mini · 2025-04-14

The current regional availability matrix lists this model in Australia East for Standard/Regional usage billing.

Watch for: This version is Deprecated (no longer available to new customers) and retires on 14 April 2027. Existing deployments can continue until then; new subscriptions should not plan on it.

gpt-4o · 2024-11-20

The current regional availability matrix lists gpt-4o 2024-11-20 in Australia East for Standard/Regional usage billing.

Watch for: This version is in Legacy status and retires on 14 April 2027 (Microsoft names gpt-5.1 as the replacement, which is not listed for Standard in Australia East). Microsoft notes that already-retired Regional Standard deployments remain retired.

text-embedding-3-large · version 1

The current regional availability matrix lists this embedding model in Australia East for Standard/Regional usage billing.

Watch for: The embedding request is regional; vector index and source-document residency remain separate configuration decisions. Retires 9 February 2028.

text-embedding-3-small · version 1

The current regional availability matrix lists this embedding model in Australia East for Standard/Regional usage billing.

Watch for: The embedding request is regional; vector index and source-document residency remain separate decisions. Retires 9 February 2028.

Gemini 3.5 Flash · gemini-3.5-flash

Google’s per-model processing table documents Australia ML processing for this model in australia-southeast1, but only through Provisioned Throughput, not pay-per-token.

Watch for: There is no documented Sydney pay-per-token route: Standard PayGo is offered only on the global endpoint (which may process anywhere) and the US/EU multi-regions. Australian processing requires reserved capacity. Retirement is listed as 19 May 2027 or later.

Gemini 3.5 Flash-Lite · gemini-3.5-flash-lite

Not available for Australian processing. Google’s per-model ML processing table marks this model for the US, EU and India only, and the model page does not list australia-southeast1.

Watch for: Corrected on 9 Nov 2026: an earlier version of this directory listed Australian processing for this model. Use Gemini 3.5 Flash with Sydney Provisioned Throughput if Australian processing is required.

DeepSeek V3.2 · deepseek.v3.2

Invoke the model in the Sydney Bedrock region through its regional endpoint. AWS lists standard on-demand token pricing.

Watch for: Use the regional model identifier and endpoint directly. Cross-region inference profiles have a different and changeable routing boundary.

Amazon Nova Lite · amazon.nova-lite-v1:0

Amazon Nova Lite supports direct in-region inference in Sydney using Bedrock on-demand access.

Watch for: Use the bare in-region model ID. Nova Lite has no Australia-only (au.) profile; its apac. profile from Sydney can route to Tokyo, Seoul, Osaka, Mumbai and Singapore.

Claude Opus 5.5 · au.anthropic.claude-opus-5-5

The au. geographic inference profile routes Claude Opus 5.5 only between Sydney (ap-southeast-2) and Melbourne (ap-southeast-4), so both destinations are in Australia.

Watch for: Use the au. profile ID, not global. or apac. Single-region Melbourne inference is also documented through the bedrock-mantle endpoint with the bare model ID. Other Claude models (Sonnet 5, Haiku 4.5, Opus 5, Sonnet 4.6) also list au. profiles; check each model card.

Qwen3 235B A22B 2507 · qwen.qwen3-235b-a22b-2507-v1:0

A large open-weight Qwen model invoked in-region in Sydney with on-demand token billing.

Watch for: AWS lists this model as In-Region in Sydney with Geo and Global profiles not supported, so the bare model ID is the route. Check account enablement and the endpoint (bedrock-runtime or bedrock-mantle) the model uses.

gpt-oss-120b · openai.gpt-oss-120b-1:0

OpenAI’s open-weight 120B model hosted by Bedrock in-region in Sydney. This is not the OpenAI API.

Watch for: AWS lists this model as In-Region in Sydney with Geo and Global profiles not supported, so the bare model ID is the route. Check account enablement and the endpoint (bedrock-runtime or bedrock-mantle) the model uses.

Mistral Large 3 · mistral.mistral-large-3-675b-instruct

Mistral’s large open-weight model invoked in-region in Sydney with on-demand token billing.

Watch for: AWS lists this model as In-Region in Sydney with Geo and Global profiles not supported, so the bare model ID is the route. Check account enablement and the endpoint (bedrock-runtime or bedrock-mantle) the model uses.

Titan Text Embeddings V2 · amazon.titan-embed-text-v2:0

An embedding model invoked in-region in Sydney. Vector store and source-document locations are separate decisions.

Watch for: AWS lists this model as In-Region in Sydney with Geo and Global profiles not supported, so the bare model ID is the route. Check account enablement and the endpoint (bedrock-runtime or bedrock-mantle) the model uses.

databricks-gpt-oss-20b · system.ai.gpt-oss-20b

The model is available to AWS Sydney workspaces on a pay-per-token endpoint, but Databricks requires cross-Geo processing for pay-per-token Foundation Model APIs outside the US and EU, so requests can be processed outside Australia and New Zealand.

Watch for: Corrected on 9 Nov 2026: Databricks’ region table shows where a model is available, not where requests are processed. Pay-per-token in the Australia–New Zealand Geo is only available with cross-Geo processing enabled.

databricks-gemma-3-12b · system.ai.gemma-3-12b

Databricks lists a hosted Gemma endpoint for AWS Sydney workspaces, but pay-per-token Foundation Model APIs in the Australia–New Zealand Geo require cross-Geo processing, so requests can leave Australia.

Watch for: Corrected on 9 Nov 2026: availability in a Sydney workspace is not processing in Sydney. Pay-per-token here is only available with cross-Geo processing enabled.

GPT-4o onshore instance; GPT-3.5 Turbo, embeddings, Claude 3.5 / Claude 4.5 Sonnet, Amazon Nova, Cohere and Stability AI with routes varying by model

A government-managed API gateway for eligible Australian Government agencies, with specific onshore model arrangements.

Watch for: The published model list does not map every model/version to its onshore route. Agency-only access and onboarding exclude general commercial self-service.

Model/version selection pending

Geography-based processing with usage billing. The exact model and deployment type matter.

Watch for: An Australian resource using Global Standard or an APAC Data Zone is not an Australia-only configuration.

Model/profile selection pending

Model access through Bedrock, with processing boundaries that depend on the invocation route.

Watch for: An Australian source region does not prove Australian inference. In-region model IDs and the au. geographic profile (Sydney + Melbourne) stay in Australia; apac. and global profiles do not.

Model/endpoint selection pending

Evaluate the model’s processing commitment separately from its endpoint location.

Watch for: A regional endpoint alone does not guarantee in-region ML processing.

Eligible models and endpoints only

Australian storage for eligible API projects, without an Australian inference processing guarantee.

Watch for: Australian storage is not Australian inference. The documented Australian region does not support regional processing.

Sydney options include databricks-gpt-oss-20b and databricks-gemma-3-12b

Pay-per-token Foundation Model APIs in the Australia–New Zealand Geo require cross-Geo processing. Provisioned-throughput and external-model endpoints have different boundaries.

Watch for: Databricks’ model region table shows availability, not processing location. Pay-per-token outside the US and EU depends on cross-Geo processing.

19 AU-region model APIs advertised; named examples include mystica_15, TF Sorcerer and TF Mystica

A smaller provider advertising Australian regional inference and OpenAI-compatible APIs.

Watch for: The provider advertises regional inference and zero retention. Its privacy policy separately allows platform/workspace data retention and overseas subprocessors.

DeepSeek V4.1 Flash and GLM-5.3; Kimi K3 listed as coming soon

A self-service, OpenAI-compatible API advertised as running in Australian data centres, with named DeepSeek and GLM models.

Watch for: Australian data-centre processing is a provider claim. The terms name a Singapore company, Xinference Holdings Pte. Ltd., as the contracting party under Singapore law, so an Australian hosting claim does not make the contract Australian.

Catalogue includes zai/glm-5.3, google/gemma-4-26b-a4b-it and qwen/qwen3-coder-30b-a3b-instruct; verify live rates

An Australian-operated, serverless API with shared inference and no dedicated capacity provisioning or monthly minimum stated.

Watch for: Australian infrastructure and prompt/output non-storage are provider statements; the public terms do not establish an Australia-only processing guarantee or fixed metadata deletion period.

Shared catalogue includes GLM-5.2, Kimi K2.6, GPT-OSS 120B, Mistral Medium 3.5, Qwen3 Coder 30B and Gemma 4 26B

A Melbourne-based Australian API provider advertising Sydney-hosted, token-metered inference without a reserved-capacity purchase.

Watch for: Sydney hosting is a provider statement, not a signed Australia-only commitment. The gateway and account database use US-headquartered vendors (Vercel and Supabase) in their Sydney regions. Each response returns an x-continuum-region header that can be logged per request.

Pilot catalogue includes gemma-4-26b-a4b-it, gpt-oss-120b, glm-5 and glm-5.2; landing-page model list differs

A pilot API with inference advertised at a Mascot (NSW) facility, an office in Richmond (VIC), an OpenAI-compatible interface and a published per-token rate table.

Watch for: This is not currently a generally available self-serve API: signup needs an access coupon, credit issuance is paused, and the site terms say it is for testing and benchmarking only, with no SLA and no production or regulated data. The site's legal text still contains placeholder entity and ABN fields.

Selected model routes are listed in Australia; processing varies by model

A local developer API with 13 models listed as processed in Australia and a broader catalogue with mixed/global routes.

Watch for: Models not on the Australia list are processed offshore, and some Australia-listed models also run in other regions (e.g. GLM-5.3: DE, SG, AU). The API has one base URL (https://api.scx.ai/v1) with no per-region routing to configure, so confirm how an Australian route is pinned.

LocAI-Go, Pro, Flash, Talk and Reader aliases; Flash is documented as Muse Glimmer 30B, other upstreams only partly disclosed

An Australian API gateway with two compatible interfaces, token-priced model aliases and prepaid billing.

Watch for: Sydney processing is a provider claim, not an Australia-only commitment. Its privacy policy allows overseas infrastructure and model providers, and some upstreams are not named publicly.

Selected models use SCX capacity; exact region depends on model and route

An Adelaide-based API layer that provides a simpler interface to inference capacity, including selected SCX-backed models.

Watch for: Kubox is an access layer, not separate underlying inference capacity. Australian processing depends on the selected upstream model and route; confirm post-trial pricing before relying on the service.

Current catalogue routes through Amazon Bedrock; region is not published

An Australian-operated API gateway that gives developers a single interface to model services, currently including Bedrock-backed routes.

Watch for: On 9 Nov 2026 the provider's site returned HTTP 502 errors, so the claims below could not be rechecked. Even when it is online, an Australian operator does not establish where Bedrock processes a request.

claude-sonnet-4-6 · CORTEX_ENABLED_CROSS_REGION = 'AWS_AU'

With the account in AWS Australia and CORTEX_ENABLED_CROSS_REGION set to AWS_AU, Cortex routes requests only to the request region, Sydney and Melbourne.

Watch for: The setting matters: AWS_APJ adds Tokyo, Seoul, Osaka, Mumbai, Hyderabad and Singapore, and ANY_REGION is global. Organisations created after 9 Mar 2026 default to ANY_REGION, and a pending behaviour-change bundle enables cross-region where the parameter is unset. This is SQL/REST inside Snowflake, not an OpenAI-compatible endpoint.

Anthropic models with an au region, e.g. anthropic/claude-sonnet-5.5 and anthropic/claude-opus-5.5

A gateway that can pin inference to an Australian data centre per request, and fails closed with HTTP 400 if it cannot honour the region.

Watch for: Only inference is pinned. Vercel's own gateway processing (including TLS termination) can occur in any Vercel region, and gateway-run tools and provider safety review are not pinned. The underlying Australian inference provider is not named. Only about seven Anthropic models list an au region.

Telnyx-hosted open models, e.g. moonshotai/Kimi-K2.6, GLM-5.3, DeepSeek V4.1 Flash, Gemma 4 26B A4B; AUS model list not published

An OpenAI-compatible API with a strict data-residency mode: requests sent to api.telnyx.com.au with region AUS run in that region or fail with HTTP 422.

Watch for: Default routing is best-effort and does not guarantee location. Strict mode covers inference only (ingress, logging and billing are separate), works only for Telnyx-hosted models, and Telnyx does not publish which models are deployed in AUS. Marketing claims go further than the documentation.