Introducing ChatGPT and Whisper APIs

## ChatGPT and Whisper APIs: The Structural Shift

### 1. The core facts and immediate context

OpenAI is opening programmatic access to `gpt-3.5-turbo` — the model powering ChatGPT — at **$0.002 per 1,000 tokens**, roughly 10× cheaper than the previously available `text-davinci-003` endpoint. Simultaneously, Whisper v2-large becomes available via API at **$0.006 per minute** of audio transcription. Before this announcement, developers wanting ChatGPT-grade conversational capabilities had to either use expensive Davinci models, maintain their own fine-tunes, or negotiate enterprise access on a case-by-case basis.

### 2. Why pricing is the most important signal

A 90% cost reduction on the conversational model isn't a marginal adjustment — it crosses several critical economic thresholds. At $0.002/1k tokens, a 500-turn conversation (~750k tokens total) costs roughly **$1.50**. That's the territory where previously uneconomical use cases — automated customer support, personalized tutors, assistants embedded in B2C freemium apps — become viable without subsidy. At ~$0.02/1k tokens, `text-davinci-003` made those same scenarios economically fragile at scale.

For Whisper, the API removes the main operational friction: until now, running Whisper in production meant managing your own GPU infrastructure (cost, latency, maintenance). At $0.006/minute, one hour of transcription costs **$0.36** — competitive against Google Speech-to-Text (~$0.064/minute standard) and well below AWS Transcribe at mid-range volumes.

### 3. Technical architecture: ChatML format and context management

The `gpt-3.5-turbo` endpoint introduces a structured message format (`system`, `user`, `assistant`) that fundamentally differs from classic prompt completion. This `ChatML` format imposes design discipline: developers must explicitly manage conversational context, inject system instructions separately from user content, and manually construct history on each call. It's more verbose than a simple prompt, but enforces a clean separation between instruction and content — reducing naive prompt injection risks and making behavior auditing easier.

The context window remains **4,096 tokens** for `gpt-3.5-turbo` at launch, identical to GPT-3. For applications requiring long conversations or large documents, context management (sliding summaries, retrieval augmentation) remains an engineering constraint the API itself doesn't solve.

### 4. Potential losers and forced repositioning

**Cohere and AI21 Labs** see their core value proposition — accessible LLMs via API at reasonable cost — directly attacked. Cohere had been pricing its `command` models comparably to Davinci; the new OpenAI pricing creates immediate pressure on their rate cards.

**Audio transcription startups** (AssemblyAI, Rev AI, Deepgram) now face a competitor combining Whisper quality (better multilingual accuracy than most alternatives on public benchmarks) with aggressive pricing. AssemblyAI charges ~$0.015/minute standard — 2.5× more expensive than Whisper API.

**Integrators who built abstractions on top of GPT-3** (LangChain, Dust, etc.) must adapt their pipelines to the new message format. Not a catastrophic breaking change, but unplanned migration work.

Finally, **Microsoft** — partner and investor — benefits directly via Azure OpenAI Service, but direct availability through api.openai.com reduces the differentiated advantage Azure could pitch to hesitant enterprise customers. The coexistence of both channels creates a go-to-market tension both parties will need to manage carefully.