OpenAI API vs ChatGPT Subscription: Billing, Limits and When to Use Each

A ChatGPT subscription and the OpenAI API are two different products with two different billing systems. A subscription gives a person access to ChatGPT in a browser or app for a flat plan price, while the API lets software call OpenAI models and charges for the tokens it uses. Paying for one does not give you the other, and choosing the wrong one is the usual reason people hit unexpected costs or blocked requests.

Who this is for and what you will have

This is for business owners, IT administrators and developers deciding how to give a team access to OpenAI models. By the end you will understand how each product is billed, how API limits and usage tiers work, how to see and control spend, and which route fits common scenarios. Plans, prices and limits change often, so use this article for the model of how things work and confirm current numbers on OpenAI's pricing and limits pages.

The core difference

OpenAI's help center describes billing in ChatGPT and the API platform as separate systems, and notes that API usage is charged by the number of tokens used after you add a payment method in your API account's billing settings. OpenAI's help articles indicate that a ChatGPT subscription can be billed through the web, the Apple App Store or Google Play, so subscriptions bought in several places can produce several separate charges.

How the API is billed

API pricing is set per one million tokens. OpenAI's pricing page splits the rates into three token types:

• Input tokens, which are the prompt, at the standard rate.
• Cached input tokens, at a steep discount when you reuse prompt content.
• Output tokens, at the highest rate.

Several things adjust the price:

• Processing tiers. Standard is the baseline. Batch is roughly half the standard rate for asynchronous work. Flex offers rates similar to batch on supported models. Faster tiers cost more, and the page says the fast tier was formerly called Priority.
• Long-context requests. Requests with more than 272K input tokens cost more than short-context ones on the models the page covers.
• Regional endpoints. Regional data-residency endpoints and FedRAMP endpoints carry a 10% uplift on eligible models.
• Built-in tools. Web search, file search and code execution are billed separately, per call, per storage amount or per session, and the tokens they use are billed at the model's normal rates.
• Other modalities. Audio, realtime, image and transcription models use their own units, such as per minute or per image.

Because the bill follows tokens, you can estimate it before you build: multiply expected input and output tokens per request by the per-million-token rates, then multiply by request volume. Use the current pricing page for numbers, and add a margin for retries and longer-than-expected outputs.

How API limits work

OpenAI's rate limits guide explains that limits cap how often you can call the API. They are measured in several ways and whichever you hit first triggers the limit:

• RPM, requests per minute.
• RPD, requests per day.
• TPM, tokens per minute.
• TPD, tokens per day.
• IPM, images per minute.

Twenty small requests can exhaust a 20-RPM limit even if you have used far fewer tokens than your TPM allowance. Limits apply to your organisation and its projects, not to individual users, and they vary by model. Some model families share a limit, and long-context requests for some models have their own separate limit.

Usage tiers

Your tier rises automatically as your total credit purchases grow. The guide lists these tiers and monthly usage limits:

Higher tiers generally allow higher RPM and TPM. To see your own limits, go to Settings > Organization > Limits in the API platform and check the Rate limits section. You can also select Upgrade tier there.

Spend limits are different from rate limits

Spend controls are separate from usage limits. The guide states that a spend alert only sends a notification, while a hard spend limit causes affected requests to return a 429 error. That means a hard limit set too low will look like an outage to your users. Set alerts well below the hard limit so you have time to react.

Handling limit errors in code

When a request fails with a rate-limit response, the documentation recommends the following:

• A 429 with the slow_down code means your request rate rose too quickly, even if you are within your RPM and TPM.
• A 503 with server_is_overloaded means the model is temporarily overloaded.
• In both cases follow the Retry-After header when present. Otherwise use exponential backoff with random jitter and cap the number of retries.
• Do not retry quota or billing errors, because those need action from you.
• Once you reach about 1 million input TPM, increase traffic by no more than 50% every 15 minutes.

Responses can include headers that show your current status, such as x-ratelimit-remaining-requests, x-ratelimit-remaining-tokens and x-ratelimit-reset-tokens. Log them so you can see how close you are to a limit before you hit it. If you run models through Azure instead, the limits and errors differ, and the same ideas are covered in fixing Azure OpenAI 429 rate limit errors.

Data handling

For the API, OpenAI's data controls guide states that, as of March 1, 2023, data sent to the OpenAI API is not used to train or improve OpenAI models, and that data is used for training only if you explicitly opt in. Abuse-monitoring logs are generated for all API feature usage and retained for up to 30 days by default, and may be kept longer where the law requires. Zero Data Retention is an approval-based control that requires prior OpenAI approval and additional requirements.

Consumer and workspace plans for ChatGPT have their own data controls and terms, which are set out on OpenAI's policy pages. If you plan to put company data into either product, read the terms for the specific plan you are buying and have your security or compliance owner approve it.

When to use each

Use a ChatGPT subscription when

• People want to chat, draft, summarise and analyse documents interactively.
• You want a predictable seat price and no development work.
• The work is ad hoc and varies by person.
• You want OpenAI's built-in app features without building an interface.

Use the API when

• You are building a product feature, internal tool or automation that calls a model.
• You need batch processing of many documents or records.
• You want control over the model, instructions, output format and logging.
• You need per-project cost tracking and limits.
• You want to connect the model to your own data or systems.

Use both when

Many organisations give staff a chat product for day-to-day use and run the API for specific workflows. Keep the two budgets separate, because they are billed separately, and put an owner on each.

Cost control checklist for the API

1. Create separate projects for separate workloads, so limits and spend are easy to read.
2. Set a spend alert at a comfortable level and a hard limit above it.
3. Estimate cost before launch from expected tokens per request and request volume.
4. Reuse long prompt prefixes to benefit from cached input pricing, and put reusable content at the start of the prompt.
5. Move non-urgent work to batch processing for the roughly 50% lower rate where the model supports it.
6. Pick the smallest model that meets the quality bar, and test before upgrading.
7. Log token usage per request so you can find expensive prompts.
8. Review usage and limits monthly, and revisit your tier as purchases grow.

For comparison with another vendor's pricing model, see Claude API pricing and how to estimate token costs.

Troubleshooting

You pay for ChatGPT but API calls fail. The subscription does not provide API access or credit. Add a payment method or credits in the API platform's billing settings and check your tier.

Requests return 429 although usage looks low. Check which limit you hit. It could be requests per minute, a short burst that triggered slow_down, a model-specific limit, or a hard spend limit.

The bill is higher than expected. Look for long prompts resent on every call, high output lengths, retries, tool charges for web or file search, and use of a faster processing tier.

You cannot find your limits. Open Settings, then Organization, then Limits in the API platform, and read the Rate limits section.

References

• OpenAI, Rate limits guide: https://developers.openai.com/api/docs/guides/rate-limits
• OpenAI, API pricing: https://developers.openai.com/api/docs/pricing
• OpenAI, Data controls in the OpenAI platform: https://developers.openai.com/api/docs/guides/your-data
• OpenAI Help Center, Billing settings in ChatGPT vs Platform: https://help.openai.com/en/articles/9039756-billing-settings-in-chatgpt-vs-platform
• OpenAI Help Center, How can I move my ChatGPT subscription to the API: https://help.openai.com/en/articles/8156019-how-can-i-move-my-chatgpt-subscription-to-the-api