What does each kind of AI project cost?
These are the ranges we quote for business clients. They are AUD, ex GST. They cover design, build, backend, testing, store submission and 3 months of free support after launch. AI usage and hosting are extra.
Other Australian agencies publish wide ranges for the same kinds of project. When you compare quotes, check what each one includes and excludes.
| Project | Examples | Cost (AUD, ex GST) | Timeline |
|---|---|---|---|
| One AI feature in an existing app | Photo or document recognition, summaries, smart search | $15k – $40k | 3 – 8 weeks |
| Internal AI tool (web) | Triage inbound email, draft quotes, extract data from PDFs | $25k – $50k | 1 – 2 months |
| AI assistant over your content | Chat that answers from your help docs, policies or product data (RAG) | $30k – $70k | 1 – 3 months |
| AI-first mobile app | A new iOS and Android app where AI is the core, like Pantry Pic | $50k – $100k | 3 – 4 months |
| AI agents and complex products | Several AI features, agents that take actions, integrations, admin tools | $100k+ | 5+ months |
In our projects the model is rarely the expensive part of the build. The cost sits in the app around it: sign-in, payments, the backend, an admin view, testing on real data and store review. For the non-AI side, see our app development cost guide.
What pushes an AI build up or down?
- Your data. Clean, structured data is cheap to use. Scanned PDFs, messy spreadsheets and data spread over several systems take much longer to prepare.
- How wrong it’s allowed to be. A slightly odd recipe idea is a small problem. A wrong insurance answer is a big one. The lower the tolerance, the more time goes into testing, guardrails and human review.
- Actions, not just answers. An assistant that only replies is simpler than one that books, refunds or edits records. Each action needs permissions, logs and an undo path.
- Speed. Showing replies word by word, or returning photo results in a second or two, needs more backend work than a job that can finish in the background.
- Offline and on-device. Running a model on the phone avoids API fees but limits which models you can use and adds testing on older devices.
- Compliance. Health, finance and anything for children need more care and paperwork. The OAIC is developing a Children’s Online Privacy Code. Some contracts and sector rules also require Australian data hosting. Note that Google’s Gemini API terms bar apps directed at, or likely to be used by, under-18s.
What do the AI models cost per request?
Most apps call a hosted model from OpenAI, Google or Anthropic. You pay per token, which is roughly three-quarters of a word. Input (what you send) and output (what comes back) are priced separately.
These are the list prices on each provider’s official pricing page on 26 October 2026. All three publish them in US dollars. Your invoice currency depends on your billing account.
| Provider | Model | Input per 1M tokens | Output per 1M tokens |
|---|---|---|---|
| OpenAI | gpt-6-astra (most capable) | US$10.00 | US$50.00 |
| OpenAI | gpt-6.1-sol | US$2.00 | US$10.00 |
| OpenAI | gpt-6-luna (cheapest) | US$0.10 | US$0.50 |
| Gemini 4 Argon | US$2.00, rising to US$4.00 on 1 Dec 2026 | US$10.00, rising to US$20.00 on 1 Dec 2026 | |
| Gemini 3.8 Flash | US$0.75, rising to US$1.50 on 1 Jan 2027 | US$3.75, rising to US$7.50 on 1 Jan 2027 | |
| Gemini 3.1 Flash-Lite | US$0.25 | US$1.50 | |
| Anthropic | Claude Fable 5.1 (most capable) | US$10.00 | US$50.00 |
| Anthropic | Claude Opus 5.5 | US$4.00 | US$20.00 |
| Anthropic | Claude Sonnet 5.5 | US$2.00 | US$10.00 |
| Anthropic | Claude Haiku 5.5 | US$0.10 | US$0.50 (prompts up to 100k tokens) |
Sources: OpenAI pricing, Google pricing and Anthropic pricing.
Note the two Google price rises. A cost model built on October prices will be wrong by January. We recheck provider prices before every quote.
What does an image cost?
Reading a photo and making a picture are priced differently. Reading is billed as input tokens. Making is billed as output, quoted per image. The table shows only the image part of each request. Your instructions and the text that comes back are billed on top.
| Task | Model | Image part only (USD) |
|---|---|---|
| Read a photo | Gemini 3.1 Flash-Lite (1,120 tokens at default resolution) | About $0.0003 |
| Read a photo | Claude Haiku 5.5 (1,296 tokens for 1000×1000 px) | About $0.00013 |
| Read a photo | Claude Opus 5.5 (1,296 tokens for 1000×1000 px) | About $0.0052 |
| Make an image | GPT Image 2, 1024×1024, low / medium / high | $0.006 / $0.053 / $0.211 |
| Make an image | Nano Banana 2 Lite (Gemini 3.1 Flash Lite Image), 1K | $0.0336 |
| Make an image | Nano Banana 2 (Gemini 3.1 Flash Image), 1K | $0.067 |
Token counts come from Google’s media resolution docs and Anthropic’s vision docs. Image prices come from OpenAI’s image guide and Google’s pricing page.
Ways to pay less
- Batch work that can wait. Anthropic and OpenAI both take 50% off batch requests.
- Cache the part of the prompt that repeats. A Claude Sonnet 5.5 cache hit costs US$0.10 per 1M tokens, against US$2.00 fresh. Writing to the cache costs US$2.50, and the default cache lasts 5 minutes.
- Resize photos on the phone before upload. For many tasks a 1000 px photo does the job of a 12-megapixel one. Test it first if small text or small objects matter.
- Route by task. Use a small model for high-volume work and a big one only where it earns its price.
- Watch the regional uplift. OpenAI adds 10% for regional processing on newer models. Anthropic adds 10% for US-only inference.
Worked example: AI cost per user per month
Take an app where people photograph something, get a short written result, and can chat with an assistant about it. Here’s the maths with the prices above. We’ve used A$1 = US$0.70, rounded from the rate on 23 October 2026.
This is model usage only. Hosting, retries, monitoring and maintenance are separate. Output estimates include any thinking tokens, which are billed as output.
- Photo: 1,720 input tokens × US$0.25 per million = US$0.00043. 400 output tokens × US$1.50 per million = US$0.0006. Total US$0.00103.
- Assistant reply: 5,000 cached tokens × US$0.10 per million = US$0.0005. 1,000 fresh tokens × US$2.00 per million = US$0.002. 500 output tokens × US$10 per million = US$0.005. Total US$0.0075.
- That assumes the 5,000 repeated tokens are already cached. The first reply in a session pays the cache write instead: 5,000 × US$2.50 per million = US$0.0125, so that reply costs US$0.0195.
- If 10 of the 40 replies pay the cache write, the month comes to US$0.47, about A$0.67. With no caching at all, every reply costs US$0.017 and the month comes to US$0.73, about A$1.05.
- A heavy user does five times as much. With a warm cache that’s US$1.76, or about A$2.50 a month.
- Add 10 generated images a month at GPT Image 2 medium and you add US$0.53. Image generation can cost more than all the text put together.
| Task | Model | Tokens per use | Cost per use (USD) | Example user a month |
|---|---|---|---|---|
| Read a photo | Gemini 3.1 Flash-Lite | 1,720 in, 400 out | $0.00103 | 30 × = $0.031 |
| Write a short result | Claude Haiku 5.5 | 2,000 in, 1,000 out | $0.0007 | 30 × = $0.021 |
| Assistant reply | Claude Sonnet 5.5, 5,000 tokens cached | 6,000 in, 500 out | $0.0075 | 40 × = $0.30 |
| Total | US$0.35, about A$0.50 |
Now check it against revenue. A subscription at A$9.99 a month is A$9.08 after GST. Apple keeps 30% in a subscriber’s first year and 15% after that. Members of its Small Business Program pay 15% from the start. At 15% you keep about A$7.72. At 30% you keep about A$6.36.
At A$7.72, the example user’s AI cost is about 6.5% of that. A heavy user is about 32%. Free users cost money too: 10,000 free users doing 5 photo reads each is 50,000 reads, or about US$52 a month.
That’s why we set free limits and price plans on the heavy user, not the average one.
What does an AI app cost in its first year?
The build is the biggest cheque, but not the only one. Here’s the build plus the first 12 months after launch for a new AI-first app. All figures are AUD, ex GST, and the maintenance and hosting lines are our budget assumptions. Your numbers will differ.
| Item | Assumption | Build plus 12 months (AUD) |
|---|---|---|
| Build | Mid-range AI-first app, ex GST | $70,000 |
| Maintenance | 15% of the build a year, after 3 months of free support: $70,000 × 15% × 9/12 | $7,875 |
| AI usage | 2,000 active users × A$0.50 × 12 months | $12,000 |
| Hosting and services | Serverless backend at A$150 a month | $1,800 |
| Apple Developer Program | Yearly fee, A$144.99 in Australia | About $145 |
| Google Play | US$25, once | About $36 |
| Total | About $91,900 |
AI usage grows with users, so it’s the line to watch. For more on ongoing costs, see the app maintenance cost guide.
App store rules for AI features
- Apple requires you to say when personal data goes to a third party, including third-party AI, and to get explicit permission first. See guideline 5.1.2(i) in the App Review Guidelines.
- Google Play requires apps covered by its AI-generated content policy to let users report or flag offensive content without leaving the app. Some limited uses, such as summarising, are out of scope. See the AI-generated content policy.
- If people can create an account, Apple requires a way to start deleting it inside the app. Google Play also wants a web page where people can request deletion, listed in Play Console.
- Plan a consent screen, a report button and both deletion flows into the first build. Adding them after a rejection costs a review round.
RAG or fine-tuning: which do you need?
For an assistant that answers from changing business documents, start with good prompts and RAG. RAG (retrieval-augmented generation) looks up the right pieces of your content at question time and hands them to the model with the question. Fine-tuning trains a model on your examples so it behaves a certain way. The two can be combined.
| RAG | Fine-tuning | |
|---|---|---|
| Best for | Answering from your documents, products or policies | A fixed style, format or narrow task done thousands of times |
| Keeping it current | Update the documents and re-index. New answers follow once indexing finishes | Usually a poor way to keep changing facts up to date |
| Shows its sources | Yes, if you build citations in. It can quote and link the passage it used | Not on its own. Pair it with retrieval if you need sources |
| What you need | Your content, cleaned and split into passages | Good example pairs, and a provider that still offers it. OpenAI’s self-serve fine-tuning is closed to new customers |
| Build cost | Part of the $30k – $70k assistant range above | Adds training runs, an evaluation set and hosting choices |
| Running cost | Embeddings are cheap, at US$0.02 per 1M tokens for OpenAI’s text-embedding-3-small. Storage, retrieval and the answer itself cost extra | Training fees, plus fine-tuned models can cost more per token |
In our experience, better prompts and RAG solve most requests for fine-tuning. We fine-tune when a smaller tuned model can match a big model on one task and save real money at volume. Check availability first: OpenAI stopped new customers creating fine-tuning jobs on 7 May 2026, and existing customers lose it on 6 January 2027. See OpenAI’s deprecations page.
What does the Australian Privacy Act require of an AI app?
If your app handles personal information, plan for the Privacy Act 1988 and the Australian Privacy Principles (APPs) from day one. This is general information, not legal advice. Check your case with a privacy lawyer.
Who it applies to
The APPs cover businesses with annual turnover over $3 million. Some smaller businesses are covered too, including health service providers and anyone who trades in personal information. The OAIC’s small business page lists them. Many clients follow the APPs anyway, because app stores, investors and enterprise customers expect it.
New from 10 December 2026: automated decisions
The Privacy and Other Legislation Amendment Act 2024 added APP 1.7 to 1.9. From 10 December 2026, your privacy policy must cover some automated decisions. It applies when a computer program makes, or substantially and directly helps make, a decision using personal information. The test is whether that decision could significantly affect someone’s rights or interests.
The policy must list the kinds of personal information used, the kinds of decisions made by the program alone, and the kinds it helps with. The OAIC released guidance and a fact sheet on 30 September 2026.
An ordinary recipe suggestion is unlikely to meet that test. Approving a loan, pricing insurance, screening a job applicant or cutting off a service could. It covers any software, not just AI.
Sending data to overseas AI providers
- Sending personal information to an overseas provider is usually a disclosure under APP 8, though some tightly controlled arrangements count as a use instead. Check the provider’s contract and subprocessors, and take the reasonable steps APP 8 requires. You can stay accountable for what they do with it. See the OAIC’s APP 8 guidelines.
- OpenAI doesn’t train on API data by default. It keeps abuse-monitoring logs for up to 30 days, with some exceptions. Its Australian data residency option stores data in Australia but doesn’t process it there. See OpenAI’s data controls.
- Google doesn’t use paid Gemini API data to improve its products. The free tier is different: Google may use it and humans may review it. We never ship a client app on the free tier. See the Gemini API terms.
- Claude on Amazon Bedrock can be routed so inference stays within Sydney and Melbourne, using Bedrock’s Australian cross-Region profiles from an Australian Region. Check which models support it before you commit.
Other rules to know
- The OAIC’s guidance on commercially available AI products says AI output that contains personal information is personal information, including made-up details about a real person.
- A statutory tort for serious invasions of privacy started on 10 June 2025. People can now sue directly for serious, intentional or reckless privacy invasions.
- Name your AI providers, and what you send them, in your privacy policy. Then complete the App Store and Google Play privacy forms to match.
- Budget for the work this creates: collection notices, consent screens, data minimisation, deletion and retention jobs, and a privacy impact assessment for higher-risk uses.
Is there an AI law in Australia?
Not a standalone one. The National AI Plan, released in December 2025, relies on existing laws such as privacy, consumer law and copyright. Mandatory AI guardrails were put on hold. As of October 2026 there’s still no general AI Act in force, but further AI and online safety reforms are being developed.
The government’s current voluntary guidance is the Guidance for AI Adoption from the National AI Centre. It replaced the Voluntary AI Safety Standard in October 2025 and cut its 10 guardrails to 6 essential practices:
- Decide who is accountable.
- Understand impacts and plan accordingly.
- Measure and manage risks.
- Share essential information.
- Test and monitor.
- Maintain human control.
It’s voluntary, but it’s a good checklist, and some enterprise and government buyers write governance rules into contracts. In practice it means a named owner and a test set you rerun before each release. It also means logs, a way to report a bad answer, and a human who can step in.
Using a chatbot doesn’t remove your Australian Consumer Law obligations. Claims about accuracy must be true, and limits should be explained. Budget for correcting wrong answers and handling complaints.
How we build an AI app
- Test the idea on real data first. We run your photos, documents or questions through two or three models in a week or two, before quoting the full build.
- Pick models on accuracy, speed and cost per use. Then work out the cost per active user, as in the example above.
- Build the AI behind your own backend. API keys never ship in the app, and you can swap models without a store release.
- Add a fallback provider, timeouts and usage limits per user.
- Keep a test set of real inputs and rerun it whenever a prompt or model changes.
- Ship, measure real usage, then tune prompts and models to bring the cost down.
We’re an Australian team in North Bondi, working Sydney hours, and we’ll sign an NDA before you share your data. If you have a brief, our free estimate form gets you a range. More on our approach is on the AI app development page.
Common questions
How much does it cost to add ChatGPT to an app?
Our estimate is $15,000–$40,000 (AUD, ex GST) to add a simple ChatGPT-style assistant to an existing app, or $30,000–$70,000 if it answers from your own content. It uses a model through an API, billed per token on top. In our worked example that is under A$1 per active user a month. We put the model behind your own backend so the API key stays secret.
Is it cheaper to use one AI provider or several?
One provider is simpler to build and test. Two reduces your dependence on one company’s outages and price changes, but each provider needs its own testing and a privacy review. We usually build for one main provider with a fallback for the features that matter most.
Can AI run on the phone instead of in the cloud?
Yes, for some tasks. Apple and Google both offer on-device models that cost nothing per request and work offline. They only run on newer phones, so you need a cloud fallback if the feature must work on older devices too. They suit focused tasks like short summaries, tagging and classification.
Do I own an app built on someone else’s AI model?
Your contract should assign the custom code, prompts and other deliverables to you, and list any third-party components and their licences. You don’t own the AI model. You pay for it per request under the provider’s terms. Rights in AI-generated output depend on those terms and on Australian IP law.
Can I claim the R&D Tax Incentive on an AI app?
Some AI development can qualify, where the outcome couldn’t be known in advance and you ran experiments to find out. Routine app building usually doesn’t. The rules are strict. Changes are proposed for income years starting on or after 1 July 2028. Talk to an R&D tax adviser before you count on it.
What does an AI app cost each month after launch?
Three things: AI usage, hosting and maintenance. AI usage grows with users and depends on the model, often under A$1 per active user a month for photo and text features. Hosting on a serverless backend can be under A$100 a month for a small app. Maintenance is usually 15–20% of the build cost a year.
Sources
- OpenAI API pricing
- Gemini Developer API pricing
- Anthropic Claude pricing
- Anthropic: vision and image token costs
- OpenAI: image generation costs
- OAIC: new resources on AI and automated decision-making transparency
- OAIC: guidance on privacy and commercially available AI products
- OAIC: small business and the Privacy Act
- OAIC: statutory tort for serious invasions of privacy
- National AI Centre: Guidance for AI Adoption
- OpenAI: data controls and residency
- Gemini API additional terms