What Meta AI Is Actually For, Now Its Model Keeps Changing Its License

The short version

Meta AI still looks like the free chatbot inside Facebook, Instagram and WhatsApp, but 2026 quietly rebuilt it. Its model closed in April, reopened in August, and Meta started charging: $7.99 to $19.99 a month for consumer power users, and $2 per million tokens for businesses running its WhatsApp agent since August 1.

On April 8, 2026, Meta closed the license on the model now running inside Facebook, Instagram, WhatsApp and Messenger for more than a billion people.

That model, Muse Spark, powers Meta AI, the assistant built into all four apps. Four months later, on August 10, the company reopened its weights and released a second, smaller model built to run on laptops instead of a data center.

In between those two dates, the assistant also picked up two new price tags: a paid consumer tier tested in three countries, and a per-token bill for the businesses running its customer-service agent. Anyone deciding whether to build around Meta AI, recommend it to a client, or budget for it this quarter needs that timeline, not the marketing page.

The model closed, then reopened four months later

Meta Superintelligence Labs, the research group Mark Zuckerberg formed to chase what he calls personal superintelligence, shipped Muse Spark as its first model on April 8. The company said only that it hoped to open-source future versions, and offered this one through a private preview to select partners, according to its own announcement. That was new. A decade of Llama being the open option in a market of closed labs paused, at least for a while, on one sentence.

The new model went on to power a search tab the company added to Facebook in mid-June, which answers questions using public posts and Groups instead of returning a list of links. It also became the model behind the assistant's voice conversations and camera-based answers across every app that carries its name.

Then, on August 10, the reversal: the company said it would open the weights on Muse Spark 1.2 after all, and released a second, separate model called Muse Glimmer, 30 billion parameters, built to run on one consumer graphics card, free to download under an open license. Four months is not a strategy. It is a company that changed its mind in public, twice, about whether its flagship assistant should run on something you can download.

Two new price tags appeared this summer

Until May, the assistant's entire business model was free. That changed on the 27th, when Meta announced a subscription ladder covering its social apps and its AI: a $7.99-a-month tier and a $19.99 tier with deeper reasoning for complex tasks. Testing began the following month in three countries, not everywhere at once.

The second price tag landed on businesses, not individuals. The assistant's customer-service agent, the version that answers people on WhatsApp, Instagram and Messenger, went from a free pilot to metered billing on August 1: $2.00 per million tokens, or roughly four to five cents per conversation, since a typical exchange runs 20,000 to 25,000 tokens.

A free test window that ran through the previous month ended on schedule. By an earnings call ten days later, Zuckerberg told investors that more than a million businesses were already using the company's agents every week.

Who this actually helps

For the roughly one billion people who never leave the free apps, not much changed. Zuckerberg told investors on that same call that daily interactions with Meta AI were up 60% since it was rebuilt around the new model, and none of that growth required anyone to pay anything.

For a small business already living on Instagram DMs, the pitch is genuinely cheap: an agent trained on your own catalog, for a few cents a message, with no seat license. That undercuts almost anything a five-person shop could build or hire for on its own.

Who this actually hurts

The businesses that lose are the ones already selling AI customer support for a living. Intercom and Decagon both price their support agents per resolution, a model built for a world where the cheapest AI option still cost real money. A free-in-app agent now competes with them on price for the same job, and it never has to win a procurement conversation because it is already installed.

The open-source community lost something less obvious: trust in the roadmap. Developers who read the April announcement and built around the closed API instead of waiting are now four months into a bet the company itself reversed.

The obvious objection

None of this touches the person who just wants a chatbot inside a messaging app. Licensing status and per-token invoices are operator-level concerns, invisible to almost all of the billion-plus people who use the assistant without ever reading a pricing page. That is true, and it will stay true for most of them.

But the direction is the story. A subscription test in a handful of markets and live billing in the world's largest messaging market are aimed at the same question: which slice of a billion free users can be converted into a paying one. ChatGPT built its business by adding a paid tier on top of a free product. This is the identical experiment, just starting from a much larger free base.

What to watch

Two dates matter next. October 1 is when billing resumes for service-window replies, the messages businesses send inside the standard 24-hour support window, a change the company has confirmed but not yet priced. And however the subscription test performs will decide whether it expands into markets where the assistant faces real competition from Perplexity and other paid tools, or stays a quiet experiment that never leaves its first three countries.

The open license on the smaller model is the one part of this that will not reverse easily. Once a 30-billion-parameter model is downloaded onto laptops under an open license, it cannot be closed back up the way the larger one was in April. Whatever the company decides about pricing its assistant, it has already given away the one thing it cannot take back.

Frequently asked questions

What is Meta AI actually for in 2026?

Meta AI is the free assistant built into Facebook, Instagram, WhatsApp and Messenger, answering questions, generating images and now powering AI Mode search on Facebook. As of August 2026 it also underpins two paid products: a Meta One consumer subscription and a metered Business Agent for companies replying to customers on WhatsApp.

Does Meta AI still run on Llama?

No. Meta AI now runs on Muse Spark, a model from Meta Superintelligence Labs that launched closed on April 8, 2026, and had its weights reopened as Muse Spark 1.2 on August 10, 2026. Llama remains Meta's older open-source line, separate from the Muse series.

How much does Meta AI cost now?

The core assistant is still free for casual use. Meta One adds paid tiers, Plus at $7.99 a month and Premium at $19.99, currently testing only in Singapore, Guatemala and Bolivia.

How much does Meta's WhatsApp Business Agent cost?

Since August 1, 2026, Meta charges $2.00 per million tokens for Business Agent conversations on WhatsApp, Instagram and Messenger. A typical customer exchange runs 20,000 to 25,000 tokens, which works out to roughly four to five cents per conversation.

Is Muse Glimmer the same as Muse Spark?

No. Muse Spark is the model that powers the Meta AI assistant across its apps. Muse Glimmer is a separate, smaller 30-billion-parameter model Meta released on August 10, 2026, built to run locally on a single consumer GPU under an open Apache 2.0 license.

Covered in this guide

Meta AI: Meta's free AI assistant, available inside Facebook, Instagram, WhatsApp, and Messenger.
ChatGPT: ChatGPT is OpenAI's AI assistant with 900 million weekly users and GPT-5.5, covering writing, coding, image generation, and web search with a free plan and Plus at $20/month.
Decagon: Decagon is a $4.5B-valued AI concierge that automates chat, voice, email and SMS support for enterprises, with contracts typically starting near $95K/year.
Intercom: Customer service platform for SaaS teams: Fin AI auto-resolves up to 80% of queries at $0.99 each. Used by 25,000+ organizations, starting at $29/seat/month.
Muse Glimmer: Muse Glimmer, released August 10, 2026 by Meta AI, is a 30B open-weight agentic model built to run on a single consumer GPU.
Muse Spark: Muse Spark, released April 8, 2026 by Meta Superintelligence Labs, is Meta's first proprietary frontier model with a 262K-token context and 1491 Elo on Chatbot Arena.
Perplexity: Perplexity AI searches the live web and returns cited, source-backed answers, used by 45 million monthly users with plans starting free.

Sources

Still deciding?

This guide covers a handful of options. Smart Match checks every listing in the directory against how you actually work and what you can spend, then hands you the shortlist and the reason behind each pick.

Start Smart Match

Related guides

All AI guides Browse the AI directory