OpenAI API logo

[ AI tools · Models and assistants ]

Product Features on the OpenAI API, Built to Last

We build assistants, extraction, search and agents on OpenAI’s GPT models, with the right tier for each step, structured outputs, evaluations on your data and costs you can predict. As an OpenAI Select Partner, this is work we do every week.

Model tier

Most product features: extraction, drafting, routing

  • Quality on hard tasks80
  • Cost efficiency75
  • Speed80

Illustrative relative scores. We measure the real ones on your own examples before choosing.

OpenAI’s Developer Platform, on One Board

The models behind ChatGPT, available to your product through an API, with controls for output format, tools, cost and data handling. Order what the feature needs, nothing more.

The API menu

Core

Responses APItext · images · tools

One API for text, images and tools, with conversation state and built-in web search, file search and code interpreter.

Choose the size

Model tiersflagship · mini · nano

Flagship, mini and nano models, plus adjustable reasoning effort, so each step pays only for the intelligence it needs.

Shape the answer

Structured outputsyour JSON schema

Responses that always match a JSON schema you define, so code can rely on the shape of the answer.

Let it act

Tools and MCPfunctions · MCP

Function calling and remote MCP servers, so the model can look things up and act in your systems.

Talk to it

Realtime voiceWebRTC · WebSocket

Speech-to-speech models over WebRTC or WebSocket for voice agents that can be interrupted.

Pay less

Batch and cachingbatch · prompt cache

Discounted batch processing for offline jobs and prompt caching for repeated context.

[ Inside one request ]

Five Steps Between an Email and a Safe Reply

A support email arrives. Our code sends it with instructions and the customer's order history to a mini model.

  • Tool call: The model asks to look up the order. Our code runs the lookup with read-only access and returns the result.
  • Structured output: The answer comes back as JSON matching our schema, so the next step never has to parse free text.
  • Checks: Rules in our code confirm the refund is within policy; low-confidence cases go to a person.
  • Logged and measured: Input, output, cost and latency are logged, and the case joins the evaluation set used before every release.

Step 01 · Request

  • input: "My blender arrived cracked..."
  • model: mini tier
  • context: last 3 orders

Example request, simplified.

[ Cost ]

The Same Feature at a Quarter of the Bill

Most of an AI feature’s running cost is a design decision. These are the moves we make on almost every build. Bars show an illustrative relative cost, not a quote.

  1. Every request on the flagship model100
  2. Route easy requests to mini, hard ones to flagship?
  3. Cache the long, repeated instructions?
  4. Move overnight jobs to the Batch API?
Engineer smiling while working at her desk
Example output01 / 06Answer with 3 cited help articlesInvoice INV-2207 to 14 validated fieldsTop 5 passages, each with its sourceRefund over limit sent for approvalCall answered in under a secondDraft reply ready for the agent to edit
[ What we build ]

OpenAI-Powered Features We Ship for Clients

  1. 01

    Customer assistants

    Answers from your help center with citations, and hand-off to people when the assistant is unsure.

  2. 02

    Document extraction

    Invoices, contracts and forms turned into structured, validated records.

  3. 03

    Search over your data

    Retrieval on your documents with answers grounded in the sources they came from.

  4. 04

    Agents that act

    Multi-step work with tools and approvals, built with the Agents SDK.

  5. 05

    Voice agents

    Phone and in-app voice on the Realtime API.

  6. 06

    Copilots in your product

    Drafting, summarizing and suggesting inside your own interface.

[ Data handling ]

What Happens to Your Data on the API

  • Not used for training

    OpenAI states that data sent through the API is not used to train its models by default.

  • Retention options

    Abuse-monitoring retention by default, with zero data retention available for eligible use cases.

  • Data residency

    Regional storage and processing options for eligible customers, including Europe.

  • Or through Azure

    The same models are available in Microsoft Foundry for companies that need Azure's controls.

Three Orders We Take Most

Packages from $1,500; custom projects from $8,000. API usage is billed by OpenAI to your account; our fees cover the build.

Order 01

AI Chatbot Starter

  • Assistant on your documents
  • Citations in every answer
  • Hand-off to a person
  • Live on your site or app

Time1 week

Total$1,500 to $2,500

Order 02

AI Agent Integration

  • One agent on your tools
  • Function calling or MCP
  • Approvals for risky steps
  • Logging and traces

Time1 to 2 weeks

Total$4,000 to $6,000

Order 03

Production AI feature

  • Model selection on your data
  • Structured outputs
  • Evaluations in CI
  • Cost controls and monitoring

TimeTypically four to eight weeks

TotalFrom $8,000

API Questions

Showing 6 of 6

  1. Q1

    Is data sent to the OpenAI API used for training?

    OpenAI states that API data is not used to train its models by default. Eligible customers can also use zero data retention and regional data residency. We document the data flow for your security review.

  2. Q2

    Which OpenAI model should we use?

    Usually several: a mini model for most steps, a flagship model for the hard cases and a nano model for high-volume tagging. We test candidates on your real examples and choose by quality, cost and latency.

  3. Q3

    How do you keep outputs reliable?

    Structured outputs that must match a schema, checks in our code for anything with consequences, and an evaluation set of your real cases run before every release.

  4. Q4

    Should we use OpenAI directly or through Microsoft Foundry?

    Directly gets new features first and is simplest to start. Through Microsoft Foundry suits companies on Azure that want Entra sign-in, private networking and Azure billing. The code changes little either way.

  5. Q5

    How do you control cost?

    By routing each step to the cheapest model that passes the evaluation, caching repeated context, batching offline work and logging cost per request with alerts.

  6. Q6

    Are we locked in to OpenAI?

    We keep prompts and model calls behind a small interface, so switching a step to another provider is a configuration and evaluation exercise, not a rewrite.

[ Contact ]

Let's build your next product together

Book a free strategy call and leave with a clear plan and estimate. No commitment.

  1. 01Pick a time that suits you
  2. 0230 minutes on scope, stack, timeline and budget
  3. 03Fixed-price proposal, NDA on request
Prefer email? contact@sajaltech.com