Gov quietly launches onshore instance of GPT-4o for APS

The federal government has quietly launched a sovereign-hosted instance of OpenAI’s GPT-4o for use by the Australian Public Service (APS).

The multimodal large language model was made available to select government agencies in recent months, according to a presentation published by the Australian Public Service Academy (APSA) [pdf].

The launch is one of several key artificial intelligence milestones reached across the Australian Public Service (APS) over the past six months, culminating in the formal commencement of a whole-of-government pilot program known as GovAI.

Led by the Department of Finance, the program aims to fast-track AI adoption within the public service, provide access to AI tools for all agencies and reduce duplication of resources and investment.

The program is currently in a closed beta phase with selected agencies but is expected to be expanded by November this year.

Since January this year, several “foundations” have been laid, including embedding a Microsoft-based AI hosting service within the APS-wide collaboration platform, GovTEAMS.

According to the presentation, several agencies have onboarded to the hosting service so far, while early scoping for Amazon-hosted AI services also commenced before the end of FY25. 

A presentation timeline published by APSA reveals that the APS has rolled out a new general-purpose AI chat app for staff and developed two prototype applications.

The first, named the Assessment Assistant app, is designed to help staff gain initial insights into applications or submissions, such as grant proposals.

The other, known as the Knowledge Assistant app, aims to demonstrate how staff can use AI to query complex datasets, such as procurement rules.

Other foundation measures include developing high-level design documents for security-assessed applications and introducing templated apps to enable faster development for small and medium-sized agencies.

The program is being supported by a GovAI working group, which brings together technical experts from across government, as well as training on models including Gemini, Claude and GPT.

It is expected to be open to new AI models and hosting options by November.