OpenAI opens up ChatGPT and Whisper APIs from $0.002

On March 1, 2023, OpenAI released the GPT-3.5 Turbo and Whisper APIs; the low entry price offers a lesson in calculating the total cost of a product dependent on a provider.

OpenAI has made the ChatGPT API available to developers today, powered by the gpt-3.5-turbo model, along with an API for Whisper, its speech recognition system. The move turns two tools that have so far had a highly visible public presence into components other companies can integrate directly into their products.

The news matters because of both the price and the simplicity of access. Instead of training its own model or maintaining complex infrastructure, an application can send a request to OpenAI’s servers and receive a conversational response or an audio transcription.

A ChatGPT model for third-party products

The gpt-3.5-turbo API costs $0.002 per 1,000 tokens, OpenAI has announced. A token is a unit of text: in English, it is roughly equivalent to four characters or three-quarters of a word, although the ratio varies by language. The company puts the price at one-tenth the cost of its previous GPT-3.5 models. Primary source

This does not mean embedding the ChatGPT website inside an application. Developers access the model through an application programming interface, or API: a standardized channel that lets one program request a task from another service. They can define the assistant’s role, provide a conversation history and send the user’s question from their own interface.

The model supports up to 4,096 tokens across the prompt and response. That limit allows for short conversations and common text tasks, but it requires developers to manage how much context is retained during longer exchanges. A company looking to summarize lengthy documents, for example, will have to split them into chunks or select the relevant information in advance. Primary source

The API’s arrival is accelerating use cases that were already being tested with language models: customer-service assistants, writing tools, search interfaces for internal documents, programming support and conversational tutors. The change does not guarantee that these products will be reliable. GPT-3.5-turbo can produce convincing but incorrect answers, so services using it in sensitive areas will need review processes, clear limits and mechanisms for escalating difficult cases to a person.

Whisper puts voice on the same bill

OpenAI has also launched an API for Whisper, its automatic speech recognition model, which it released as open-source software in September 2022. The service costs $0.006 per minute of audio and can transcribe recordings and translate audio from other languages into English. Primary source

Until now, integrating Whisper meant downloading and running the model on a company’s own servers—an option useful for organizations that need technical control or process large volumes, but less accessible to small teams. The API removes that installation step: an interview platform, meeting tool or news outlet can send an audio file and receive text ready for processing.

Speech recognition is not the same as understanding a conversation. Proper names, background noise, multiple speakers and unusual pronunciation remain sources of errors. For publishing subtitles, legal transcripts or medical records, automatic transcription should still be treated as a first draft, not a final document.

Low price, high dependence

The lower cost could change which projects are viable. At $0.002 per 1,000 tokens, many conversational features stop being an expensive experiment and acquire a low marginal cost. That benefits small businesses and independent developers, but it also concentrates an essential part of their product in OpenAI’s infrastructure and commercial terms. Primary source

The company has said it will retain data sent to the API for 30 days to monitor potential abuse and will not use it to improve its models unless the customer gives explicit consent. That is an important condition for organizations looking to analyze internal information, although it does not eliminate the need to assess what data can be sent to an external provider. Primary source

The next challenge will be distinguishing between a superficial integration and a useful product. The API makes it easier to add a chat box or automatic transcription; designing safeguards, protecting data and taking responsibility for model failures will remain the job of whoever builds the service.

Per-token price is only one line on the bill

An application pays for inputs and outputs, but also storage, retrieval, moderation, retries, observability and human review. Calculate monthly cost from a real distribution of conversations: average and extreme tokens, failed responses and auxiliary calls. Multiplying the price by one short demo usually underestimates production.

The context window creates another choice. Keeping the entire history raises cost and can carry irrelevant information; summarising saves tokens but loses detail. A robust system keeps source data outside the model, selects necessary passages and makes it possible to reconstruct why an answer received that context.

Whisper adds different units: duration, audio quality, language, latency and correction of proper names. Cheap transcription can be expensive when someone must repair every sentence. Take a representative sample of calls or recordings, create a human reference and measure error by type rather than through general impression.

Dependency and reversibility

The launch post priced GPT-3.5 Turbo at $0.002 per thousand tokens and offered Whisper through a managed API. Those prices were dated. A product should tolerate changes in tariffs, limits, versions and availability. Budgets and contracts cannot assume the introductory number will remain.

Design exit before dependency. Preserve an internal format for messages and transcripts, abstract the provider call and test a minimal alternative. There is no need to switch every month; the team needs to know which data and functions would prevent it doing so if necessary.

Regression evaluations are also required. When the model changes, rerun a task set and compare accuracy, tone, safety, latency and cost. An update may improve the average while breaking an instruction the product considered essential.

The transferable skill is to calculate cost per accepted result and record dependency. Add every call and review minute, test proprietary audio and text, and document how to replace the provider. An API lowers the entry barrier; engineering determines whether it lowers total cost.

Measure latency through percentiles rather than an average alone. A response that is usually fast may stall an interface under load and long requests. Set a maximum time, decide what the user sees and avoid automatically retrying a call that could duplicate an action.

Logs should not copy every model input indefinitely. Debugging needs version, timings, size, outcome and an identifier; sensitive content should be minimised or protected. Observability does not mean creating a second database containing every conversation.

Set a spending ceiling and alert before production. An unexpectedly long input, retry loop or abuse can multiply calls without adding value. Limits per user and task contain both cost and harm; exceptions should be logged and require an owner.

Finally, budget for exit: exporting data, adapting prompts, rerunning evaluations and training the team. The reserve seems unnecessary while everything works, but prevents a price increase or version withdrawal turning technical dependency into a business emergency.