TL;DR
Organisations use OpenAI in three ways: ChatGPT for their people, the API to build their own tools and agents, and OpenAI models built into platforms they already run, such as Salesforce, Atlassian and Microsoft 365. Models change every few months, so choose by testing against your own tasks, not by version number. In Australia, check what "data residency" covers: OpenAI can store data here, but processing currently happens offshore unless you use specific Microsoft Foundry deployments, which offer older models. The value comes from connecting AI to your own systems and data, with governance that meets the Privacy Act.
We first wrote about OpenAI when GPT-5 launched in August 2025. Since then OpenAI has released several model generations, and GPT-5 itself is being retired from the API in December 2026. Rather than chase each release, this guide covers the decisions that stay the same, and keeps the version detail in one section we update.
Three ways organisations use OpenAI
- ChatGPT for your people. A managed ChatGPT workspace for drafting, analysis, research and, increasingly, multi-step work across files and apps.
- The API, to build your own. Your developers or partners use OpenAI models in your own applications and agents, either directly from OpenAI or through Microsoft Foundry on Azure.
- Inside the platforms you already run. Salesforce Agentforce, Atlassian Rovo, Microsoft 365 Copilot and Adobe's tools all use OpenAI models among others. You often get OpenAI without buying it directly.
Most organisations end up with all three. The governance question is the same across them: what data goes where, who can use it, and who is accountable for the outcome.
ChatGPT Business or Enterprise?
| ChatGPT Business | ChatGPT Enterprise | |
|---|---|---|
| Price | US$20 per user per month billed annually, or US$25 monthly | By quote |
| Sign-in | Single sign-on (SAML) and multi-factor authentication | Adds SCIM user provisioning and identity groups |
| Your data | Not used for training by default | Adds data retention controls, data residency, and encryption keys you manage |
| Oversight | Workspace admin | Adds a Compliance API for audit logs, usage and credit controls, and custom roles |
| Suits | Smaller teams getting started | Larger or regulated organisations |
For regulated Australian organisations, the Enterprise controls usually decide it: SCIM to remove leavers automatically, audit logs for compliance teams, and retention settings you can match to your records policy.
ChatGPT can also search connected apps such as SharePoint, Google Drive and Slack, respecting each user's existing permissions. Connected apps are on by default in Business and off by default in Enterprise, so decide which ones to allow before rollout.
The current OpenAI models
OpenAI's flagship API models are:
| Model | Best for | Input / output |
|---|---|---|
| GPT-6 Astra | The most demanding reasoning and analysis | $10 / $50 |
| GPT-6.1 Sol | Complex work at a lower cost; near-Astra performance | $2 / $10 |
| GPT-6 Luna | Focused, high-volume tasks such as classification and extraction | $0.10 / $0.50 |
All three handle very long inputs, over a million tokens. OpenAI also publishes open-weight models, gpt-oss-120b and gpt-oss-20b, under the Apache 2.0 licence, for organisations that need to run a model on their own infrastructure.
Retirements to plan for. According to OpenAI's deprecations page, GPT-5 and o3 shut down in the API on 11 December 2026, and GPT-4o, GPT-4o mini and GPT-4.1 on 14 January 2027. The Assistants API was switched off on 26 August 2026. If anything you built still calls these, it needs migrating and retesting now.
How to choose a model
- Test on your own work. Build a set of 30 to 100 real examples with known good answers, and run each candidate model against them. Benchmarks rarely match your documents, your terminology or your edge cases.
- Start cheap, step up only where needed. A small model often handles routing, extraction and summaries. Reserve the expensive models for the steps that need them.
- Design for change. Keep prompts, tests and model choice in configuration, so moving to the next model is a test run, not a rebuild.
- Count the whole cost. Token prices are only part of it. Integration, monitoring and the people who review outputs usually cost more.
Data residency in Australia: what it covers
This is the question we're asked most, and the answer has two parts: where data is stored, and where it's processed.
- OpenAI API: the Australian endpoint stores data in Australia, but processing isn't kept in the region. It also requires OpenAI's approval for reduced abuse monitoring or zero data retention.
- ChatGPT Enterprise: OpenAI added Australia as a data residency region in November 2025 for data at rest. In-region processing for ChatGPT is currently offered only in the US and Europe.
- Microsoft Foundry (Azure): deployment type decides where prompts are processed. In Australia East, regional deployments keep processing in Australia but offer older models: GPT-4o and GPT-4.1 mini on pay-as-you-go, or up to GPT-5.6 Sol with reserved (provisioned) capacity. The newest GPT-6 models are available only on global deployments, which may process anywhere, or, for GPT-6.1 Sol, on Data Zone deployments processed within Asia Pacific.
In practice, decide per use case. Content that's public or low-sensitivity can use the newest models on global deployments. Personal, health or financial information may justify an in-region deployment and an older model, or keeping that data out of prompts altogether.
Building with OpenAI
- Responses API: OpenAI recommends it for all new projects. It handles tools, files and multi-step work. Chat Completions is still supported.
- Agents: the open-source Agents SDK covers agents, hand-offs between them, guardrails and tracing. A managed Agents API entered public beta in September 2026. OpenAI's visual Agent Builder shuts down on 30 November 2026, so don't start anything new on it.
- Connecting to your systems: the API can call remote MCP (Model Context Protocol) servers, so an agent can use tools from Jira, Salesforce or your own systems through a standard interface, with approval required by default. See our MCP integration service.
- Privacy: API data isn't used for training unless you opt in, and abuse-monitoring logs are kept for up to 30 days unless OpenAI approves zero data retention.
OpenAI in the platforms you already run
- Salesforce: OpenAI is a preferred model provider for Agentforce, and customers can make it the default model for Agentforce's reasoning engine and Prompt Builder.
- Atlassian: Rovo uses OpenAI models alongside models from Anthropic, Google and open-source providers, and prompts aren't used to improve them. A Rovo connector also brings Jira and Confluence into ChatGPT.
- Microsoft: Microsoft 365 Copilot runs on OpenAI models, with Anthropic models also available. In Copilot Studio, used to build agents for Microsoft 365 and Dynamics 365, the default OpenAI models run in-region in Australia.
- Adobe: Firefly offers OpenAI's image models, and Adobe released a plugin bringing Photoshop, Acrobat, Express and other tools into ChatGPT in August 2026.
Each platform's AI works well inside that platform. The gaps are between them: an agent in Salesforce can't see the Jira work behind a customer escalation, and Copilot can't see the order history in your commerce platform, unless someone connects them. That integration work, with permissions and data handling designed in, is where most of the value sits.
Governance and Australian law
- Guidance for AI Adoption: the National AI Centre's six essential practices (October 2025): decide who is accountable, understand impacts, measure and manage risks, share essential information, test and monitor, and maintain human control.
- No standalone AI Act: the National AI Plan (December 2025) relies on existing, largely technology-neutral laws, so privacy, consumer and anti-discrimination law all apply to AI.
- Privacy Act, automated decisions: from 10 December 2026, privacy policies must explain when computer programs make or substantially make decisions that significantly affect people. See our guide to the automated decision rules.
- OAIC guidance on AI products: the regulator's guidance on commercially available AI recommends not entering personal or sensitive information into public AI tools, doing privacy impact assessments, and checking accuracy.
A workable policy covers approved tools, what data may go into each, who reviews outputs before they're relied on, and how incidents are reported.
Getting value, not just usage
Adoption is no longer the problem. In McKinsey's State of AI in 2026 survey, eight in ten respondents said AI had improved their own productivity, but enterprise-level financial impact hasn't kept pace. The organisations that close that gap pick a few processes, connect AI to the systems and data those processes run on, and measure the result in business terms.
How Great Minds helps
We help Australian organisations choose where OpenAI fits, set up ChatGPT and API environments with the right data controls, and connect AI to Salesforce, Atlassian, Microsoft and the rest of the stack. Our agentic AI systems integration work builds agents that act across platforms, with permissions, logging and human review designed in.
Frequently asked questions
Is ChatGPT Enterprise data used to train OpenAI's models?
No. Business and Enterprise data isn't used for training by default, and the same applies to the API unless you opt in.
Can we keep OpenAI data in Australia?
Storage, yes, with the OpenAI API's Australian endpoint or ChatGPT Enterprise data residency. Processing in Australia currently requires a regional Microsoft Foundry deployment in Australia East, which offers older models.
Should we use OpenAI directly or through Microsoft Foundry?
Foundry suits organisations already on Azure that want Azure's security, billing and regional options. OpenAI directly gets the newest models and features first. Many organisations use both.
Which model should we use?
Test the current models against a set of your own real tasks. Use the smallest model that passes, and keep the choice in configuration so you can switch as models change.
What happened to GPT-5?
It has been succeeded by newer models and shuts down in the API on 11 December 2026. Applications built on it should be migrated and retested before then.
Where to start
List the AI tools already in use across your organisation, approved or not, and the data going into each. Then pick one process where AI could save real time if it could see data from more than one system. That's usually the best first project, and we're happy to help scope it.