ChatGPT integration through OpenAI’s API, with the key, the bill and the answers undercontrol.
A quick test with ChatGPT is easy. A feature your customers rely on is a different job. That is where our ChatGPT integration services come in. We connect the model to your app, website or internal tool through your own server. We choose the model for the task, test the prompts on your real examples, make the output something your code can rely on, and set a limit on what it can spend.
Your app talks to OpenAI. What happens on a bad day?
Calling the API takes a few lines of code. The trouble comes later: a key left where anyone can read it, answers that change between runs, and data kept longer than the owner knew. Published research and OpenAI's own pages show how common each one is.
3,000
live websites, roughly, were found with an OpenAI API key readable in their public page code
What we doWe keep your OpenAI key on your server, so the page your visitors load never contains it.
Five problems owners bring to us after a first try with the OpenAI API, and what we do about each.
The problem
Our developer put the OpenAI key in the website code, and now I am told anyone can copy it.
What we do
We move every call behind your own server. The browser or the mobile app asks your server, and your server asks OpenAI. OpenAI's documentation says the same: do not expose a key in client-side code. The exposed key is replaced, and the new one sits in a project of its own.
The key5 done for you
Key held on your server only
Exposed keys revoked and replaced
A separate project for the live app
Test and live kept apart
Sign-in required before a call is made
The problem
The OpenAI bill jumped one month and nobody could tell me why.
What we do
We set a spend limit and an alert on the project in your OpenAI account, and a cap in your app on what one user or one request can use. Each call is logged with what it used, so a jump can be traced to a feature, a user or a fault.
The bill5 done for you
A spend limit on the project
An alert before the limit is reached
A cap per user and per request
Usage recorded for every call
Long inputs trimmed before they are sent
The problem
The same question gets a different answer each time, and sometimes it breaks the page.
What we do
We ask the model for structured output, which holds a completed answer to a format your code defines. Your code still checks each answer, including the case where the model declines. We test the prompts on your real examples, more than once, before launch.
Answers that vary5 done for you
Output held to a set format
Every answer checked by your code
Declined and cut-off answers handled
Prompts tested on real examples
Repeat runs compared before launch
The problem
We got an email saying the model our app uses is being switched off, and I do not know what that breaks.
What we do
We keep the model name in one setting, so a change is one edit and not a search through the code. We watch OpenAI's deprecations page for shutdown dates. You keep the test set, so the replacement is run on the same examples before you switch.
Retired models5 done for you
Model name held in one setting
Shutdown dates watched
A replacement tested on your examples
Results compared side by side
The switch made before the deadline
The problem
When OpenAI is slow or down, our feature just spins. And I never agreed what customer data gets sent.
What we do
We handle the errors OpenAI returns for rate limits and overload by waiting and trying again, with a limit on tries. When a call still fails, users see a message or a simpler path we agreed with you. We also list every field sent to OpenAI and leave out what the task does not need.
Limits, outages and data5 done for you
Retries that wait longer each time
A time limit on every call
A fallback your users can use
A list of the data that is sent
OpenAI's data terms read and dated
Three bad days nobody planned for.
Three public events, taken from OpenAI's own pages and from press reports, show what building on one provider's API involves. None of them is our client.
01
Outage · December 2024
One internal change took the API down for over four hours
What happened
OpenAI's own write-up says its API, ChatGPT and Sora began to degrade at 3:16 PM Pacific time on 11 December 2024 and recovered at 7:38 PM. The cause was a new monitoring service that overloaded the systems running its servers. OpenAI said it was not a security incident.
What we do about it
We put a time limit and a fallback on every call, so your users see a clear message or a simpler path when the API is down.
A court told OpenAI to keep data it would have deleted
What happened
VentureBeat reported that a US court order of 13 May 2025, in the New York Times case, required OpenAI to preserve output logs that would otherwise be deleted. It reported that API customers without a zero data retention agreement were covered, and that OpenAI told users on 5 June. OpenAI objected to the order.
What we do about it
We send OpenAI only the fields the task needs and give you a dated note of its terms, so you know what could be held.
OpenAI's deprecations page lists its Assistants API with a notice date of 26 August 2025 and a shutdown date of 26 August 2026. It names the Responses API and the Conversations API as the replacement. Apps built on the older interface had one year to move.
What we do about it
We keep all OpenAI calls in one layer on your server and watch the deprecations page, so a retirement is a planned change.
Live websites we built. Open them and judge the work for yourself. They are not OpenAI integrations: our AI work was for private clients and is shared on a call, with the client's agreement.
Anyone can make the first call. The work is everything around it.
We test every AI system on real examples before it goes live, and we cap what it can spend. We've built apps with an AI recommender, AI essay feedback, an AI chatbot and AI-made language courses. We are not an OpenAI partner or reseller, so we check its documentation on the day and tell you when a different model would suit the job. We don't promise the model will always be right. We build so that a wrong answer, a slow call or a retired model is handled.
The Sun Software team
01
The key never reaches the browser
Every call goes through your server, on a project of its own in your account.
02
Tested on your examples
Prompts are scored on real cases from your business, and you keep the test set.
03
Limits set before launch
A spend limit, an alert and per-user caps are in place on day one.
04
Checked against OpenAI's own pages
Data terms, limits and retirements are read at the source and dated in your notes.
Great....very professional, delivered on time and exactly as I needed. Exceptional. I always find Sunsoft very easy to work with, always helpful and I'll continue using their services for the foreseeable future. No need to go anywhere else.
Client in United KingdomOctober 2026 · Fiverr review
“
Excellent experience from start to finish. Communication was great, responses were always prompt, and the work was completed quickly. They captured the exact look and feel I wanted on the first try and delivered excellent value for the money. I would gladly work with them again and highly recommend their services.
Client in GermanySeptember 2026 · Fiverr review
“
Choosing SunSoft was exactly the right decision. Their professionalism, extensive experience, and attention to detail have really impressed me so far. It was a complex project that exceeded my expectations! I look forward to continuing to work with SunSoft!
Client in CzechiaAugust 2026 · Fiverr review
Questions before you build on OpenAI
What do ChatGPT integration services include?
Connecting OpenAI’s models to your app, website or internal tool through your own server. We choose the model for each job and test it on your real examples, keep the API key on the server, set up structured output, add a spend limit and a fallback for failed calls, and keep a log of calls.
Why can’t my website call OpenAI directly?
Because the key would be visible to anyone who opens the page’s code. OpenAI’s own documentation says not to expose a key in client-side code such as browsers or apps, and to load it on the server. So your site calls your server, and your server calls OpenAI.
Does OpenAI train its models on what we send?
On the day we wrote this, OpenAI’s data page said data sent to its API is not used to train or improve its models unless you opt in. It also said logs kept to monitor abuse may hold prompts and responses for up to 30 days. Terms change, so we read the current page with you and note the date. This is not legal advice.
Will the same question always get the same answer?
No, and we do not promise it will. We narrow the variation with fixed instructions, structured output and settings that favour consistency. We also keep a test set, so you can measure how much the answers move after any change.
What happens when OpenAI retires the model we use?
OpenAI lists retirements on its deprecations page with a shutdown date. We keep the model name in one setting, and you keep the test set. A replacement can be run on the same examples before you switch.
Can you stop the bill from running away?
We set a spend limit and an alert on the project in your OpenAI account, and a cap on how much one user or one request can use. If the limit is reached the feature pauses and shows a message we agreed with you. It does not keep spending.
Are you an OpenAI partner?
No. We have no partnership or reseller arrangement with OpenAI or any other AI provider. We work with its API because it suits some jobs, and we say so when another model or a simpler tool would suit yours better.
What decides the cost of an OpenAI integration?
You get a written proposal once we know what the model has to do inside your app. OpenAI charges for usage separately, on your own account, under a limit we set with you. Three things decide it. The jobs: How many separate tasks the model does, and how many real examples each one has to be tested on. What it can reach: Text in and text out only, or also looking things up and taking actions in your own systems. The safeguards: How much logging, checking, limiting and fallback the task calls for, and what your contracts require.
What happens after I contact you?
We reply within one business day. Then the work runs in 5 steps: Define the job, Test models and prompts, Build the server layer, Add limits and a fallback, Launch and hand over. Each step ends with something you can review.
1Talk Through Your Integration
2We reply within one business day
3Discovery call
4Step one: Define the job
What should the model do in your app?
Tell us what you want the model to do and send a few real examples. We reply with how we would build it, what would leave your systems and what decides the cost.