ChatGPT Subscription vs OpenAI API: Separate Bills, Credits, and Keys

Plans buy app access; the API buys tokens

The grid shows what each bill covers. The last row matters as much as the price when the work involves company or client data.

ChatGPT credits are not API credits. On Plus and Pro they extend the plan's Codex and Work usage after a limit, and they stay inside ChatGPT. In the other direction, an API key does not unlock plan features: Codex signed in with a key has no Codex cloud, no ChatGPT Work on the web, no connectors and no voice.

Illustrative break-even: $20 of Plus against per-token rates

OpenAI's Codex pricing page warns against using API prices to estimate what a plan includes, and the two buy different things: Plus comes with the apps, the API with the model alone. The arithmetic still shows how far $20 goes on the API. Assume each request sends 20,000 input tokens (the prompt plus any earlier conversation or files your client includes) and returns 1,000 output tokens. These sizes are illustrative, not OpenAI figures.

Read the last column as the monthly use at which Plus becomes the cheaper route, as long as the work fits within Plus's usage limits. Twenty GPT-6.1 Sol requests a working day for 20 days reaches 400, and GPT-6 Sol, which Plus includes, has the same uncached rates. Plus also includes GPT-6 Astra, where $20 of API use runs out after 80 requests. For GPT-6 Luna work the comparison is Go, at $8, which the API matches at 3,200 requests. Someone who sends a few requests a week through their own client may spend less on the API, but gives up ChatGPT Work, deep research and the other app features. Cached input lowers the API side: GPT-6.1 Sol bills it at $0.10 per million, a twentieth of the input rate.

Software that calls the models through the API is billed only on the API, and no plan covers it. An illustrative script that classifies 50,000 short items, each 500 input and 50 output tokens, uses 25 million input and 2.5 million output tokens. On GPT-6 Luna that is $2.50 plus $1.25, or $3.75 at Standard rates, and $1.875 through Batch, which costs half and completes within 24 hours.

When to pay for a plan, the API or both

• You work in ChatGPT most days: a plan. ChatGPT Go vs Plus shows what the extra $12 adds, and ChatGPT Plus vs Pro the next step.
• Your product, script or nightly job calls a model: the API. Batch and Flex cost half of Standard, so work that can wait a day costs less.
• A team handles company or client data: Business for people working in the apps, from two users, with no training on business data by default. An API organization for code: API data is not used for training unless you opt in, abuse-monitoring logs are kept for up to 30 days by default, and eligible approved customers can use Zero Data Retention. ChatGPT Plus vs Business covers the seat decision.
• You do both: budget them separately. Paying for one never reduces the other.

Where the bills meet: Codex sign-in

Signed in with ChatGPT, Codex draws on the plan's usage, which it shares with ChatGPT Work. Signed in with an API key, every turn is billed to the Platform account at standard API rates instead of plan credits. When a plan hits its limit, any user can switch to a key and keep working locally, but the cloud features (Codex cloud, GitHub code review, Slack) need ChatGPT sign-in. Codex: ChatGPT login vs API key has the sign-in details, Codex credits vs API billing compares the two ways to keep going, and Codex CLI cost covers the command line.