OpenAI API Integration for Website and Mobile App Development
OpenAI API integration means wiring a product's real data, permissions and interface into OpenAI's models through the Responses API, and keeping that integration working as the models underneath it are replaced. Mixcore Studio builds these integrations into websites and mobile apps, and takes over ones that were built two or three model generations ago and have stopped being maintainable.
The hard part of this work in 2026 is not calling the API. It is choosing a model tier you can afford at your traffic volume, getting structured, machine-readable output you can trust in application code, and absorbing the fact that OpenAI's surface turns over roughly once a year. A page that recommended GPT-4 and the Assistants API eighteen months ago was recommending two things that now have shutdown dates.
What changed in the OpenAI API, and what is being switched off
Anyone maintaining an OpenAI integration should have these dates in a calendar rather than discovering them through a 404.
- The Assistants API shuts down on 26 August 2026 — deprecation was announced a year earlier, and the replacement is the Responses API together with the Conversations API. Threads, runs and assistant objects do not carry across unchanged, so this is a rewrite of the orchestration layer, not a base-URL swap.
- The Responses API is now the supported starting point — Chat Completions has not been shut down and existing code on it keeps working, but it is no longer what OpenAI recommends new builds start from.
- GPT-5.6 launched publicly on 9 July 2026 in three tiers — sol for long autonomous tasks, terra as the balanced default, luna for high-volume classification, summarisation and drafting.
- Legacy models have hard shutdown dates — gpt-3.5-turbo, gpt-4, gpt-4-turbo and o1 stop serving on 23 October 2026; gpt-3.5-turbo-instruct on 28 September 2026; gpt-5, gpt-5-mini, gpt-5-nano and the o3 snapshot on 11 December 2026; gpt-realtime, gpt-audio and gpt-4o-audio on 20 January 2027.
- Tooling is going too — the OpenAI Evals platform, Agent Builder and the Reusable Prompts API all shut down on 30 November 2026, so evaluation and prompt management need to live in your own repository rather than in a vendor console.
What we build on the OpenAI API
- Assistants inside existing products — a conversational surface that knows which record, order or document the user currently has open, built on the Responses API with the Conversations API holding state.
- Document and content pipelines — extraction, classification, summarisation and translation running as background jobs against gpt-5.6-luna, where the per-token price makes high volume viable.
- Search and retrieval features — embeddings with text-embedding-3-small or 3-large over your own corpus, with the model used to answer rather than to remember.
- Voice interfaces — gpt-realtime-2.1 for low-latency speech, with gpt-realtime-2.1-mini as the cost-controlled tier for longer or higher-volume sessions.
- Image generation and editing — gpt-image-2 inside product workflows, with generated assets marked as synthetic where the law now requires it.
- Migrations — Assistants API to Responses, legacy JSON mode to Structured Outputs, and retiring models to their current equivalents.
Picking a model tier, and what it actually costs
Model choice is a budget decision before it is a quality decision. OpenAI's published list prices per million tokens in August 2026 are $5.00 input and $30.00 output for gpt-5.6-sol, $2.00 and $12.00 for gpt-5.6-terra, and $0.20 and $1.20 for gpt-5.6-luna. Per OpenAI's API changelog, luna was cut by 80% and terra by 20% on 30 July 2026, so the gap between tiers is now wide enough that routing decisions dominate the bill.
Voice deserves its own modelling. At $32 input and $64 output per million audio tokens for gpt-realtime-2.1, an always-listening feature is expensive in a way text is not, and the mini tier at $10 and $20 exists for exactly that reason. We model cost per session against your expected usage before writing the feature, not after the first invoice.
In practice most products end up mixed. Cheap tiers handle the volume, the flagship handles the small number of genuinely hard requests, and a router with cached responses sits in front of both.
Making the output reliable enough to build on
Application code needs a guaranteed shape, not prose that usually parses. We use Structured Outputs with a json_schema and strict mode set to true, which enforces schema adherence including required keys and valid enum values, surfaces model refusals in a dedicated refusal field, and supports streaming partial structured data. Schemas are defined once in Zod or Pydantic and shared between the model call and the application, so a schema change cannot silently drift from the code that consumes it.
The legacy JSON mode only guarantees syntactically valid JSON, not that the fields you need are present, and OpenAI now recommends Structured Outputs wherever the model supports it. If your integration still parses free text with regular expressions or retries until the JSON looks right, that is the first thing worth replacing.
Alongside that we keep an evaluation set in your repository, run it against every model or prompt change, and instrument token spend per user, tenant and feature from the first release.
Disclosure rules that took effect on 2 August 2026
Article 50 of the EU AI Act became applicable on 2 August 2026. It requires that people are told when they are interacting with an AI system, that synthetic audio, image, video and text is marked in a machine-readable form, and that anyone exposed to emotion recognition or biometric categorisation is notified. Systems already on the market have until 2 December 2026 to meet the marking requirement.
Separately, the high-risk obligations many teams budgeted for this August have moved. Regulation (EU) 2026/1744, the Digital Omnibus on AI, was published in the Official Journal on 24 July 2026 and entered into force on 27 July 2026, pushing Annex III standalone high-risk obligations from 2 August 2026 to 2 December 2027 and Annex I embedded systems to 2 August 2028. The prohibitions in Article 5 have been in force since 2 February 2025 and were not deferred. For a typical chat or content feature this means disclosure and content marking are live work now, and the heavier conformity assessment work is not.
When the OpenAI API is the wrong choice
It is the wrong tool for anything deterministic. If the answer can come from a database query, a rules table or a regular expression, a model call is a slower, more expensive and less predictable version of code you already know how to write. We have removed model calls from products more than once for this reason.
It is also the wrong choice when data residency or procurement rules forbid sending content to a third-party US API, when you need a stable model that behaves identically in three years, or when a narrow classification task at very high volume would be cheaper and more accurate as a small fine-tuned model you host yourself. And if your organisation has no appetite for a migration roughly every twelve months, that is a real constraint worth designing around rather than discovering later. Where any of these apply we will say so and scope open-weight or self-hosted options instead.
Where it fits well is open-ended language and multimodal work over your own data, at a quality level that is hard to reach any other way, provided someone owns the upgrade path.
Our expertise
- Responses API migration
- Model tier selection
- Structured Outputs
- Token spend control
- Realtime voice integration
- AI disclosure compliance
Frequently asked questions
Our product runs on the OpenAI Assistants API. What do we need to do?
The Assistants API shuts down on 26 August 2026, so the integration has to move to the Responses API with the Conversations API holding conversation state. Threads, runs and assistant objects have no direct equivalent, so plan for a rewrite of the orchestration layer with your tool definitions and prompts carried across. We run this as a like-for-like migration first, with behaviour compared against the old implementation before any changes are made.
Which GPT-5.6 tier should we be using?
Start from the workload rather than the benchmark. gpt-5.6-luna at $0.20 input and $1.20 output per million tokens handles classification, summarisation and drafting at volume, gpt-5.6-terra at $2.00 and $12.00 is the balanced default for user-facing conversation, and gpt-5.6-sol at $5.00 and $30.00 is worth reserving for long autonomous tasks. Most products route across two tiers rather than standardising on one.
What does an always-on voice feature actually cost to run?
gpt-realtime-2.1 is priced at $32 input and $64 output per million audio tokens, which makes continuous listening expensive at scale. gpt-realtime-2.1-mini at $10 and $20 covers most production cases. We model cost per session against your expected duration and concurrency before the feature is built, and design push-to-talk or activity-gated capture where continuous audio is not justified.
Who is responsible when OpenAI retires the model our product depends on?
This should be written into the engagement rather than assumed. Legacy models currently carry published shutdown dates through to January 2027, and the pattern has been roughly annual. We keep an evaluation set in your repository so a model swap can be tested rather than guessed at, and we agree up front whether migration sits inside a support arrangement or is quoted as separate work.
Do we legally have to tell users they are talking to an AI?
In the EU, yes. Article 50 of the EU AI Act became applicable on 2 August 2026 and requires disclosure that a person is interacting with an AI system, machine-readable marking of synthetic audio, image, video and text, and notice to anyone exposed to emotion recognition or biometric categorisation. Systems already on the market have until 2 December 2026 for the marking requirement. We build the disclosure and content-marking layer as part of the integration rather than as a later retrofit.
Contacts
We are always happy to talk with you.
Feel free to contact us in any suitable way
Request a quote
Let's discuss your project!
Please, provide us with a brief description of what you
already have and what you are going to achieve.
Mail us contact@brainiacminds.com