Vol. 21 taught you to name the product: chatbots, coding assistants, voice, education, and the rest. This lecture opens Vol. 22 by answering a different question: which vendor supplies the substrate? OpenAI is the default first pick many teams evaluate—not because it is always the right answer, but because its product surface is the widest closed stack engineers already know: GPT chat, Images / DALL·E, Whisper, coding (Codex lineage), Sora video, and a metered API.
Model families lived in Vol. 11.5 GPT family and closed-source models. Modalities lived in Vol. 16 (DALL·E, Whisper, Sora). The client lived in Vol. 18 OpenAI SDK and Tier 1 API-only. Here you compare OpenAI as a procurement choice against the rest of Module 22.1. Vol. 21 HITL, eval, and privacy still apply; a GPT API is a component, not a finished product.
Learning Objectives
By the end of this lesson, students should be able to:
- Map OpenAI’s public product surfaces (GPT, Images/DALL·E, Whisper, Codex/coding, Sora, Platform API) without inventing unpublished internals.
- Use a qualitative Pricing / Strengths / Weaknesses table to decide when OpenAI is the default vs a multi-vendor or open-weight path.
- Separate ChatGPT (consumer/workspace product) from the Platform API (your backend).
- Cross-link Vol. 11.5 families, Vol. 16 modality catalogs, and Vol. 18 SDKs when evaluating a SKU.
- Treat model IDs as changing SKUs—pin in config, re-read current docs before a spike.
- Explain why Vol. 21 product patterns (HITL, eval, privacy) still sit on top of any OpenAI call.
OpenAI is a closed-weight frontier lab whose commercial stack is delivered mainly as ChatGPT (consumer and workspace apps) and the OpenAI Platform API. Public product lines include GPT language/multimodal assistants, Images (DALL·E / GPT Image lineage), Whisper speech-to-text (open weights and hosted audio API), Codex / coding agents (legacy 2021 Codex API → deprecated; current Codex relaunched in 2025 as an agentic cloud coding product alongside ChatGPT coding SKUs), and Sora video generation. Weights for flagship GPT/Sora/Image models are not generally downloadable. This volume catalogs the vendor; it does not claim secret architecture.
Product Surfaces (What You Actually Buy)
| Surface | Job | Curriculum home |
|---|---|---|
| GPT (chat / reasoning SKUs) | Text + tools + multimodal chat; flagship vs mini/nano tiers | 11.5 GPT family; Vol. 21 chat/agents |
| Images / DALL·E | Closed still generation via ChatGPT or Images API | 16.3 DALL·E |
| Whisper | ASR: open checkpoints or Audio API | 16.2 Whisper |
| Codex / coding | Code completion and agentic coding products (IDs churn) | 21.1 Coding assistants; Vol. 22.6 IDEs |
| Sora | Closed text-to-video (playground + evolving Videos API) | 16.4 Sora; Vol. 22.3 video vendors |
| Platform API + SDK | Metered HTTP: chat/Responses, embeddings, images, audio, tools | 18.1 OpenAI SDK |
ChatGPT is a product (UI, memory, GPTs, browsing). The Platform API is a contract your backend calls. Do not confuse a Plus/Team subscription with production token billing, data-retention settings, or SLA.
Pricing / Strengths / Weaknesses
Do not memorize invented dollar prices or fake leaderboard scores. Re-read OpenAI’s current pricing page and your own evals. Qualitative posture:
| Axis | Pricing (qualitative) | Strengths | Weaknesses |
|---|---|---|---|
| GPT text / tools | Token-metered API; flagship SKUs cost more than mini/nano. ChatGPT is a subscription, not the same bill. Reasoning/“thinking” modes spend extra tokens. | Broadest closed ecosystem; tool calling; Azure OpenAI / compatible gateways; fastest “everyone already has a key” path. | Closed weights; vendor lock-in; SKU churn; data-path and ToS diligence required; not automatically cheapest. |
| Images / DALL·E | Per-image / quality / size tiers; usually more expensive per asset than a short text completion. | Strong prompt following on later SKUs; one vendor for chat + stills; ChatGPT for exploration. | Closed; safety filters + latency jitter; not a self-host ControlNet studio (see Vol. 16.3 SD/FLUX and Module 22.2). |
| Whisper | Hosted audio is usage-metered; open Whisper is “you pay GPU/CPU,” not OpenAI tokens. | Excellent batch ASR default; open weights exist; simple Audio API. | Not a native streaming/diarization/PII stack—see Vol. 16.2 Deepgram et al. for live captions. |
| Codex / coding | Bundled in ChatGPT/Codex products and/or API coding SKUs; seat vs token mix depends on surface. | Deep coding narrative; pairs with Vol. 21 HITL diff review. | Historical 2021 Codex API is deprecated; 2025 Codex is an agentic coding product. GitHub Copilot is a Microsoft surface (later lecture + 22.6). |
| Sora | Clip-length / resolution / audio tiers; video generations typically dwarf text token cost. | Cinematic T2V in the ChatGPT ecosystem; evolving API. | Closed; duration/control limits; safety + watermarking; compare Veo/Runway on Vol. 16.4 rubric—no fake FPS crowns. |
| Enterprise path | Azure OpenAI and OpenAI business plans: commitments, residency, support—not a public sticker. | Procurement familiarity; SSO/logging stories via cloud partners. | Still closed weights; you do not own the model; dual-run (OpenAI + Azure) can confuse data maps. |
Pick OpenAI when
- You need one vendor for chat + tools + images + ASR quickly
- Team already ships the OpenAI SDK
- Azure OpenAI is a compliance/procurement requirement
- Vol. 21 product is chat/support/coding and you want a known API
Look elsewhere when
Always still do
- Vol. 21 HITL + Vol. 19 eval on your tasks
- Vol. 20 privacy: what is sent to OpenAI vs Azure vs local
- Pin model IDs; never treat “GPT” as one frozen checkpoint
Why teams start here
- Widest closed multimodal + tools surface
- SDK and Azure compatibility reduce time-to-spike
- Whisper open weights give an on-prem ASR escape hatch
Why teams add a second vendor
- Failover, price competition, and capability niches
- Open weights for private bulk inference
- Your FastAPI should not import only one SDK forever (Vol. 18.2)
Practical API Sketch
Teaching stand-in. Pin the current model ID from docs. Keys stay in the environment. This is a vendor call, not a product.
Related Lectures
| Lecture | Role |
|---|---|
| Vol. 21 Education AI | Previous volume: product categories before vendors |
| GPT family / Closed-source models | Model lineage vs this procurement catalog |
| DALL·E / Whisper / Sora | Modality catalogs (stills, ASR, video) |
| OpenAI SDK / Tier 1 API-only | How you actually call and size the stack |
| Anthropic | Next provider: Claude / Claude Code / API |
| GitHub Copilot | Microsoft coding surface that historically used OpenAI models |
“OpenAI” is one model. It is a catalog of SKUs and products. Second: ChatGPT Plus is the same as the Platform API (billing, retention, and SLAs differ). Third: published blog scores replace your Vol. 19 eval. Fourth: Whisper open weights mean GPT weights are open (they are not). Fifth: Codex still means the 2021 API—names reuse; re-read current docs. Sixth: picking OpenAI removes the need for HITL, privacy maps, or a second vendor.
Knowledge Check
- Short Answer: What question does Vol. 22 add after Vol. 21’s product categories? Answer: Which vendor/ecosystem supplies the substrate (APIs, models, tools)—not a new product type.
- True/False: ChatGPT Plus billing is identical to Platform API token billing. Answer: False.
- Multiple Choice: Flagship GPT weights are generally: (a) closed / API or ChatGPT access, (b) OSI open source downloads, (c) only on-device. Answer: (a).
- Short Answer: Name one OpenAI modality cataloged in Vol. 16. Answer: DALL·E / Images, Whisper, or Sora (any one).
- True/False: This lecture invents dollar prices and fake benchmarks you should memorize. Answer: False.
- Multiple Choice: Whisper is special because: (a) open weights exist alongside a hosted Audio API, (b) it is only a video model, (c) it replaces HITL. Answer: (a).
- Short Answer: Which Vol. 18 lecture is the primary OpenAI client? Answer: OpenAI SDK (18.1).
- True/False: Model IDs should be treated as changing SKUs pinned in config. Answer: True.
- Multiple Choice: A good reason to look past OpenAI is: (a) must self-host weights or use Google/Anthropic estate, (b) you dislike documentation, (c) eval is illegal. Answer: (a).
- Short Answer: Which provider lecture comes next in Module 22.1? Answer: Anthropic (Claude, Claude Code, Artifacts, Projects, API).
Key Takeaways
- Vol. 22 starts here: after naming the Vol. 21 product, pick a vendor. OpenAI is the widest closed stack, not the only answer.
- Surfaces: GPT, Images/DALL·E, Whisper, Codex/coding, Sora, Platform API—ChatGPT ≠ API bill.
- Pricing / strengths / weaknesses stay qualitative; re-read live price pages and run your own evals.
- Cross-link Vol. 11.5 families, Vol. 16 modalities, Vol. 18 SDK/Tier 1. HITL and privacy still sit on top.
- Next: Anthropic as the second frontier closed provider.
Lab: Students fill a one-page vendor card for a Vol. 21 product (e.g. support chatbot or coding assistant): OpenAI surfaces used, qualitative cost posture, data path (ChatGPT vs API vs Azure), three non-goals (no fake $ prices, no unpublished internals, no skipping HITL). Spike one chat completion with a pinned SKU.
Whiteboard: Vol. 21 product box → OpenAI catalog (six surfaces) → arrows to Vol. 11.5 / 16 / 18. Then an empty box for “second vendor” filled next lecture with Claude.
Recap: OpenAI opens the ecosystem catalog: GPT, Images/DALL·E, Whisper, Codex, Sora, and the Platform API as procurement choices after Vol. 21. Next is Anthropic.