OpenAI Platform

OpenAI Platform is the API product: you create a project, add a payment method in the dashboard, and your application calls documented endpoints. The GPT family is the headline, but the same account can expose image generation, audio (text-to-speech and speech-to-text), embeddings, batch jobs, fine-tuning where OpenAI still documents it for your account, and tool-using interfaces — Assistants or the newer Responses API, depending on what the docs show when you build. Check the current reference rather than a slide from last year.

This is not ChatGPT in a browser with a monthly seat. ChatGPT Plus is a separate consumer product. Platform usage is metered in the API dashboard. A marketer at a Bangkok SME can use ChatGPT for one-off drafts; the Platform is what you choose when a website, a CRM helper, or a nightly catalog rewrite needs a key that lives on a server.

Keys belong on a server. Pasting a secret into a landing-page script, a no-code “custom JS” box, or a shared Notion page is how accounts get drained and how guest data leaks. WordsThatSells treats the Platform as something a developer or a trusted automation host calls. The marketer still owns the brief, the approved facts, and the publish decision.

Thai and Lao quality is not guaranteed by the logo. Evaluate the specific model you intend to ship: tone marks, hotel names in Vientiane, Thai product titles that should not be “helpfully” spaced. Run that eval before production, not after a client sees a broken glyph on mobile. Vietnamese operators in HCMC should do the same for diacritics and local brand spelling.

Usage policies live on openai.com, not in this review. If your workflow includes medical claims, political ads, or scraping inboxes, read those policies yourself. We will not invent a permission that the vendor did not grant.

• Text models in the GPT family for chat, classification, and structured outputs as currently listed
• Image endpoints for generation (and related image tools the docs still publish)
• Audio: speech-to-text and text-to-speech for voice notes, IVR experiments, or course explainers
• Embeddings for search and clustering, plus a Batch API for jobs that can wait
• Tool-calling surfaces — Assistants and/or the Responses API — plus fine-tuning eligibility that you must confirm in-account

OpenAI retires models and renames product surfaces. Fine-tuning access in particular has changed for some accounts. Build against the live docs and the dashboard, not against a directory paragraph.

A common WordsThatSells pattern is a fact sheet plus a server function. The marketer maintains a short document: room types, visa notes you are willing to publish, the English name of the Lao dish on the menu. The function sends that sheet and a task (FAQ draft, alt text, email snippet). The model never becomes the source of prices. If a number is missing, the output should say so rather than guess. That discipline matters more in tourism and education than a prettier system prompt.

Speech endpoints earn their keep when English-language course sellers or hotel training teams already record WhatsApp voice notes. Speech-to-text can turn a manager’s walkthrough into a draft SOP; text-to-speech can narrate a lesson the team already wrote. Neither step should skip a listen-through. Accents from Isan, Lao, or southern Vietnam can confuse a model that was not evaluated on your files. Keep a two-minute sample set and replay it whenever you change models.

Embeddings are the quiet win for support. If you already have English help articles — refund windows, pickup points, how to add LINE — embed those pages and retrieve them before the model answers. That is different from letting the model invent a pickup point on Thanon Samsenthai. Retrieval does not require a science team; it does require clean URLs and a developer who will refresh the index when you edit the site.

Batch jobs fit catalog work: a hundred English titles that need a consistent voice, or alt text for a HCMC F&B shoot. Because batch is designed to wait, it is a poor fit for a live chat on a landing page and a good fit for a Tuesday night export. Download the result, diff it, and reject rows that added adjectives you do not use (“luxurious,” “hidden gem”) unless the brand actually talks that way.

Tool-calling (Assistants or Responses, as documented) can look like magic in a demo: the model decides to look up a booking, then replies. In production you still define the tools. A tool that hits your real reservation system needs authentication and a narrow schema. Do not expose a generic “browse the web” tool on a client site and hope it stays on-brand. If you only need a scripted FAQ, a retrieval bot is simpler to explain to the hotel owner.

Fine-tuning is not the first lever for an SME. Most tone problems die when you put the fact sheet in the prompt and add three approved examples. Fine-tuning, where still offered to your project, means datasets, evals, and a plan for when the base model is retired. Ask whether you are solving a style issue or a missing-facts issue. Missing facts do not get fixed by training on last year’s emails.

Compliance is local even when the vendor is not. Thailand’s PDPA and Vietnam’s data rules care about what you send abroad. If a form collects guest IDs, do not pipe the raw upload into an API call. Mask, minimize, and say so in the privacy policy you already publish. Usage policies on openai.com are a second gate: they can forbid use cases your client casually requested in a briefing. Read both before you quote a build.

English-language operators sometimes assume “the API is better at English, so we skip QA.” Skip that assumption. Brand voice in hospitality English from Laos is not Silicon Valley English. Keep a reviewer who knows whether “temple visit” should name a specific wat. The Platform accelerates drafts. It does not sign the page.

The Platform is pay-as-you-go in the API dashboard. We do not reprint unit prices or estimate monthly spend. Open the billing pages on platform.openai.com for the models you actually call. Set usage limits in the project so a leaked key or a loop cannot surprise the finance lead.