ChatGPT is not the API: what you actually need
ChatGPT is OpenAI’s consumer app. A ChatGPT Plus, Team or similar subscription does not include API credit, and it cannot be used as a key for an OpenAI chatbot widget. What you need is an account on the OpenAI API platform (platform.openai.com) with its own billing, and a secret key created there.
That separation is also the good news. API usage is pay-as-you-go from a prepaid balance, so you pay for the tokens your visitors actually use and nothing for seats. Our Connect OpenAI guide walks through the platform screens, and the AI keys overview explains how SiteTrellis stores and tests the key.
If a ChatGPT subscription is all you have today, the shortest path is to sign in to the API platform with the same email, add a small prepaid balance and create a project key. Nothing in the app subscription carries over.
Which OpenAI model to try, with estimated costs
OpenAI offers three models in our catalog. All three have ids and prices confirmed from OpenAI’s own model pages on 2026-10-07, so none is flagged as unverified here. Treat the figures as estimates; the OpenAI dashboard is authoritative.
| Model | Input, per 1M tokens | Output, per 1M tokens | Estimate per 1,000 conversations |
|---|---|---|---|
| GPT-6 Astra | $10.00 | $50.00 | $179 |
| GPT-6.1 Sol | $2.00 | $10.00 | $35.88 |
| GPT-6 Luna | $0.10 | $0.50 | $1.79 |
Estimates, not quotes: list prices from our model catalog in US dollars per million tokens, for 6 visitor messages per conversation, 1,500 tokens of prompt and knowledge sent every turn, and 180-token replies, with no prompt caching. Your provider's bill is authoritative.
The small model, GPT-6 Luna, is estimated at about $1.79 per 1,000 conversations, GPT-6.1 Sol at about $35.88 and GPT-6 Astra at about $179. For a support or lead-capture agent that answers from pasted facts, start with the small model and compare it with the middle one in the playground using identical questions.
One cost detail is specific to reasoning-capable models: hidden thinking tokens are billed as output. Our catalog asks for a low reasoning effort on the larger two and none on the small one, which keeps chat replies quick, but watch the usage page after your first real day. Some OpenAI models also need a verified organization or a higher usage tier, in which case you see an access error and should pick another model or complete verification.
Set up an OpenAI chatbot widget in eight steps
Open an API platform account
Sign in at platform.openai.com and create or select an organization. This is separate from your ChatGPT login settings.
Add prepaid credit
Under billing, buy a modest balance. With zero balance, requests fail with a quota error even when the key is valid. Enable auto-recharge only together with the limits in step 6.
Create a project
Make a dedicated project for the website chat. A project lets you restrict which models the key may call and track its spend apart from other work.
Create a project key
Create a new secret key inside that project, name it “SiteTrellis production” and copy it immediately. It starts with
sk-and is shown once.Add the key in SiteTrellis
Under AI providers choose OpenAI, paste the key and wait for the automatic test to show ok.
Set a project budget
In limits, set a monthly budget and an email alert threshold below it, and allow only the models you plan to use.
Choose the model and test
In the agent’s model section pick this key and a model from the list, then run your real questions in the playground.
Publish
Publish the agent and add the snippet to your site; the publishing guide covers versions and rollback.
The 429 trap: out of money, or going too fast?
OpenAI uses the same HTTP status, 429, for two very different problems, and the error text is what tells them apart.
| Message | Meaning | Fix |
|---|---|---|
| “You exceeded your current quota” | No credit left, or a project budget was reached. | Add credit or raise the budget, then press Test on the key. |
| “Rate limit reached” | Too many requests or tokens per minute for your usage tier. | Wait, let the tier grow, or add a fallback step. |
A hard project budget has a trade-off worth planning for. When it is reached, your chat starts failing until you raise it, which is exactly why a second provider behind OpenAI is worth having. Other errors you may meet: 401 (wrong or truncated key), 403 (the project does not allow that model), 404 (retired or unavailable model) and 5xx (a temporary OpenAI problem). See troubleshooting for the rest.
Cost notes and an example fallback chain
- Prepaid credit plus a project budget gives two ceilings. Set the alert threshold below the budget so you hear about it before the chat stops.
- A shorter history window and tighter knowledge text cut input tokens on every turn.
- The cost guide and the calculator let you test your own conversation lengths.
| Step | Key and model | Why |
|---|---|---|
| 1 | OpenAI project key, GPT-6 Luna | Low cost and quick for everyday questions. |
| 2 | Gemini key on a paid project, Gemini 3.8 Flash | A different company, so an OpenAI quota stop or outage does not end the chat. Gemini prices in our catalog are unverified. |
A second OpenAI key does not protect you from an OpenAI outage, although it does help if one project’s budget is exhausted. Details of what triggers the move to the next step are in the fallback guide.
Data handling and honest limits
Visitor messages, your prompt and your knowledge go to OpenAI under your organization’s terms. OpenAI documents how API data is retained and whether it is used for training, and offers data controls to eligible organizations. Read the current policy and match your privacy policy to it.
This setup does not give you a crawler (knowledge is pasted text), streaming replies, voice or WhatsApp. It also does not remove the usual model risk of confident wrong answers, so use strict knowledge mode and a handoff path. If you would rather try another provider first, compare the Claude and OpenRouter pages, or read the BYOK pillar.