ChatGPT WhatsApp Number: Why It Stopped Working and What to Use Instead

The ChatGPT number on WhatsApp stopped working because Meta rewrote its WhatsApp Business Solution Terms to prohibit AI providers from distributing general purpose assistants through the platform. The rule took effect for new API accounts on October 15, 2025 and for every existing account on January 15, 2026. Business chatbots were not banned and still run normally.

That distinction is the whole story, and most of the coverage missed it. Headlines in early 2026 said WhatsApp was banning AI chatbots. If you run a store, a clinic or a support desk and you were planning to put an AI assistant on your business number, nothing in that change stops you. What stopped was something narrower and more specific.

What Meta actually changed

The revised terms add a prohibition aimed at one category of company. The operative language reads:

AI Providers are strictly prohibited from accessing or using the WhatsApp Business Solution ... for the purposes of providing, delivering, offering, selling, or otherwise making available such technologies when such technologies are the primary ... functionality

Everything turns on the last clause. The rule applies when the AI is the primary functionality being made available. OpenAI's ChatGPT number qualified, because the assistant was the product. So did Perplexity, Luzia and Poke, all of which wound down their WhatsApp presence around the same time.

A furniture retailer whose bot answers "where is my sofa" does not qualify. In that case the AI is a component of customer service, and customer service is exactly what the WhatsApp Business Platform was built to carry. Meta's own framing has been consistent on this point: the Business API is for companies to communicate with their customers, not a distribution channel for standalone assistants.

The dates that matter

• October 15, 2025. Every business registering for the WhatsApp Business API from this date came under the revised terms immediately.
• January 15, 2026. The prohibition reached accounts that already existed. The consumer facing assistants shut down around here.
• August 1, 2026. Meta Business Agent went live, billing by token at $2.00 per million tokens.
• October 1, 2026. Non template replies inside an open 24 hour window become chargeable, whether a person or a bot wrote them.

Read in sequence, the pattern is hard to miss. Meta closed the platform to companies whose product is an assistant, then shipped its own metered AI layer nine months later, then set a date when the replies that layer generates start carrying a messaging charge.

Can I use ChatGPT on WhatsApp?

Not as a hosted assistant you message directly on WhatsApp. That route is closed and is not coming back. You can still call OpenAI, Anthropic or any other model from your own server and use it to write the replies that go out from your own business number. The model never touches WhatsApp; your application does, and the AI is simply how your application decides what to say.

This is not a loophole. It is the ordinary architecture the platform expects, and it is what every serious support automation has always done. The difference between the permitted and prohibited pattern is not the technology, it is who the conversation is with and what is being sold.

Can businesses still use AI chatbots on WhatsApp?

Yes. Structured business tasks are unaffected: order and shipping answers, support triage, appointment booking and rescheduling, one time passcodes and account verification, and lead qualification. If you can describe what your bot does in a sentence that mentions your own products, orders or accounts, you are on the safe side of the line.

The design that gets uncomfortable is the open domain assistant. A bot that will answer anything about anything, even one a business runs for its own customers, is the shape the rule targets. The practical risk for a normal company is low, but the safe build is to scope the bot to your own domain and refuse everything else with a clean handoff to a human.

How to run your own AI on your WhatsApp number

Four things have to be true before a model can answer anything.

You need a number on the Cloud API. The free WhatsApp Business app has no API at all, so no external software can read or send on it. The On Premises API was retired on October 23, 2025, which leaves the Cloud API as the only route. The number must be yours, must have a country and area code, must be able to receive a call or an SMS, and must not already be registered to another WhatsApp account.

You need a webhook that behaves. Subscribe to the messages field. Your endpoint has to return HTTP 200, because Meta retries with decreasing frequency for up to seven days against an endpoint that fails. Self signed certificates are not supported. Verify the X-Hub-Signature-256 header as an HMAC SHA256 of the raw request body rather than the parsed JSON, and remember that the payload nests entry and changes as arrays, which is where a surprising number of first integrations break.

You need templates approved in advance. Replies inside an open window are free form, but anything that starts a conversation needs a pre approved template in the marketing, utility or authentication category. Get the utility and authentication ones you need approved before launch. Marketing templates have not delivered to US phone numbers since April 1, 2025, so US outbound from a bot means utility and authentication content.

You need the bot to know your business. A model with no grounding will invent a return policy. Most of the real work is not prompt writing, it is getting your help center, product pages and policy documents into a clean, machine readable form the bot can actually retrieve from, and wiring up live lookups against your order and account systems so the answers come from your database rather than from the model's imagination.

Build the handoff to a human before you build the clever answers. The handoff is what stops a bad reply from becoming a block, and blocks feed directly into your quality rating, which in turn sets how much you are allowed to send.

What an AI chatbot on WhatsApp costs

Two meters run at once. There is the model, billed by tokens, and there is the message, billed by Meta. Until October 1, 2026 the second meter reads zero for anything your bot says inside an open 24 hour window, which is why most published cost estimates look cheap.

After that date, in window replies become chargeable at the utility rate with no free volume tier. That change lands harder on bots than on people. A human agent might send six messages to close a thread. A bot that acknowledges, clarifies, answers, confirms and follows up sends five for every conversation rather than only the ones that reach a person. The fix is not fewer conversations, it is fewer bubbles per conversation. Writing one well formed answer instead of five short ones also sidesteps the per recipient throttle, which is roughly one message every six seconds to the same user.

Meta does not publish US dollar per message rates in its developer documentation, so any cent figure attributed to Meta is really a reseller's number. Twilio's published US list price for a delivered utility or authentication template is $0.0084, made up of a $0.005 Twilio fee and a $0.0034 Meta pass through. Meta has committed to publishing final service message rates during September 2026.

Does Meta have its own AI on WhatsApp?

Yes. Meta Business Agent went live on August 1, 2026 and bills by token at $2.00 per million tokens rather than per message or per conversation. That pricing model changes how you budget: cost scales with how much context you send on every turn, not with how many customers you talk to. A verbose system prompt replayed across a long thread can cost more than the conversation is worth, which is a failure mode that per message pricing never had.

What to do next

If you were hoping to message a general assistant on WhatsApp, that option is gone and no vendor can restore it. If you are a business, the more useful question is which structured jobs on your number are worth automating first. Order status and support triage are usually the highest volume and the lowest risk, because the answers come from systems you already control.

Our full breakdown of the rules, the permitted use cases and the 2026 cost picture is on the WhatsApp AI chatbot page, which includes a checker for whether a given bot design falls inside the AI Provider rule. For the plumbing underneath it, see what the WhatsApp Cloud API does, and for the token math specifically, Meta Business Agent pricing. The October billing change is covered in detail under WhatsApp service messages.