Meta AI Explained: Llama, Muse Spark and the Assistant in WhatsApp, Instagram and Facebook
If you have tried to work out what Meta AI actually is, you have probably run into a confusing picture. The name turns up in five different places — a standalone app, a website, WhatsApp, Instagram, Facebook and Messenger — and the articles that explain it disagree with each other about which model powers it, whether it costs anything, and whether you can turn it off. Most of that confusion has a simple cause: the thing described in most explainers is out of date. Meta AI launched as a Llama-powered assistant, and for a while that description was true. It is not true any more. On 8 April 2026 Meta replaced Llama as the engine of the assistant with a new model, Muse Spark, and Meta’s own blog calls Llama 4 Maverick “our previous model”. A reader who finishes an article believing Meta AI runs on Llama 4 has been misled, and this guide says so plainly in its first section rather than burying the correction halfway down. Checked 14 November 2026 against Meta’s own newsroom, AI blog, model card, licence and developer pricing pages. One limit of this guide is stated up front, too: Meta’s meta.ai website itself could not be opened while this guide was researched, so the current assistant experience is documented here from Meta’s newsroom and AI blog posts, not from a first-hand read of the live site. Where a fact comes from a secondary source rather than a Meta page, this guide says so.
The short version, before the detail, is this. “Meta AI” is three different things sharing one name, plus a fourth layer that most explainers miss entirely. First, Llama: a family of open-weight models for developers, which no longer powers the assistant. Second, Muse: the model family that does power the assistant today, led by Muse Spark, with Muse Wonder as the newer frontier model and Muse Ember as the cheap, high-volume option. Third, the Meta AI assistant itself: the chatbot you talk to in the Meta AI app, on meta.ai, and inside WhatsApp, Instagram, Facebook, Messenger and Meta’s AI glasses. Fourth, Muse: a separate personal AI agent, launched in September 2026, that does tasks for you rather than just answering questions. On top of all four sits a paid layer, Meta One, launched in September 2026, which buys more media generation with Meta AI and more in-app features — while the core assistant stays free. This guide maps all four layers, then takes each Meta app in turn, then deals honestly with privacy and price, so that by the end you can say which “Meta AI” you mean in any given sentence — and what you can and cannot control about it.

The Short Answer: Meta AI Is Three Things with One Name
<img src="

” alt=”A person standing at a junction with three blank signposts pointing in different directions, warm daylight, minimal composition.”>
Start with the map, because every confusion about Meta AI comes from collapsing it into one thing. The name “Meta AI” is used for an open-weight model family, for a model family that powers an assistant, for the assistant itself, and — in the newest usage — for a subscription service that sits on top of the assistant. Those are four different products with different owners, different prices and different rules, and the rest of this guide keeps them apart on purpose. Whenever a sentence in this guide says “Meta AI costs…” or “Meta AI can…”, it says which layer it means. That discipline is the whole value of the article.
Layer one is Llama. Llama is a family of open-weight models that Meta publishes for developers to download and run themselves. The current generation is Llama 4, released on 5 April 2025 in two models, Scout and Maverick. Llama is real, it is still published, and it matters a great deal to developers — but it is not what answers you when you talk to the assistant. When the standalone Meta AI app launched on 29 April 2025, it was, in Meta’s own words, “built with Llama 4”. That was true for about a year. It stopped being true on 8 April 2026, when Muse Spark launched and took over the assistant at meta.ai and in the Meta AI app. If a guide, a tutorial or a roundup still describes Meta AI as “Llama-powered”, it is describing the assistant as it stood in April 2025, not November 2026.
Layer two is Muse. Muse is Meta’s current model family, built by Meta Superintelligence Labs, and it is the engine behind the assistant today. The first Muse model, Muse Spark, launched on 8 April 2026 and is what powers Meta AI at meta.ai and in the Meta AI app. Muse Wonder and Wonder Full, announced on 12 October 2026, are the most capable models Meta has built to date, aimed at agentic work. Muse Ember is the cheapest of the three on Meta’s developer API. Muse Image generates images, Muse Voice Transcribe turns speech into text, and Muse Code brings the models into a developer’s terminal. Developers who want Meta’s current models through Meta’s own API get Muse, not Llama — the Meta Model API pricing page lists only Muse models. That single fact tells you where Meta’s centre of gravity has moved.
Layer three is the Meta AI assistant: the product most people actually meet. It is the chatbot in the Meta AI app, on meta.ai, and inside WhatsApp, Instagram, Facebook, Messenger and Ray-Ban Meta glasses. It chats, it talks, it remembers things you ask it to remember, it generates and edits images, it searches the web, and it has a short-form AI video feed called Vibes. It is free to use at its core, and Meta’s subscription announcement says so in plain words: “The core experience across our apps and Meta AI has always been free, and that’s not changing.” What has changed is that there is now a paid ceiling above the free experience, in the form of Meta One.
Layer four is Muse, the personal AI agent, launched on 8 September 2026. Muse is not a chatbot and not a model — it is an agent that does things: sending an email, booking travel, filling out a form, negotiating a bill, turning a saved recipe into a grocery list. It works after you close the app, it runs in its own secure virtual machine, and it asks permission before it spends your money or sends something on your behalf. It is powered by Muse Spark today and will be powered by Muse Wonder next, starting with subscribers. It is the newest layer, the least understood, and the one that changes the question from “what can Meta’s AI answer?” to “what can Meta’s AI do?” — which is a different question, and it gets its own section later in this guide.
Two more facts belong in the short answer because they are the two questions people ask most. First, the paid layer: Meta One, launched on 15 September 2026, costs $7.99 a month for the Core bundle and $19.99 a month for Premium, and buys more media generation with Meta AI and more features across Instagram, Facebook and WhatsApp. Meta publishes no numeric usage limits for the free tier or for either paid tier, and this guide states no numbers of its own — any roundup that quotes a specific message, image or token cap for Meta AI is quoting a figure that does not appear on any Meta page read for this guide. Second, the off switch: there is no single off switch for Meta AI in WhatsApp, Instagram, Facebook or Messenger. You can mute the chat, delete what it stored, restrict what it can do in a particular conversation, and opt out of individual features — and this guide lists those controls app by app — but the assistant itself stays in the app. That is the honest headline, and the privacy section of this guide does not soften it.
The Map in One Table: Llama, Muse, Meta AI, Muse

The table below is the whole map in one place. If you read nothing else in this guide, read this table and the short answer above it: they are the same map in two forms, and everything after them is the evidence.
| Layer | What it is | Who it is for | Where you meet it |
|---|---|---|---|
| Llama | Open-weight model family; current generation is Llama 4 (Scout and Maverick) | Developers and researchers who download models and run them themselves | llama.com downloads and Hugging Face; your own servers or a cloud host — not through Meta’s own API |
| Muse (models) | Meta’s current model family from Meta Superintelligence Labs: Spark, Wonder, Ember, Image, Voice Transcribe, Code | Developers building on Meta’s models, and — through the assistant — everyone | Meta Model API (dev.meta.ai), Muse Code, and behind the Meta AI assistant and the Muse agent |
| Meta AI (assistant) | The general-purpose AI assistant: chat, voice, memory, image generation, Vibes video, web search | Anyone with a Meta account or one of Meta’s apps | Meta AI app, meta.ai, WhatsApp, Instagram, Facebook, Messenger, Ray-Ban Meta glasses |
| Muse (agent) | A personal AI agent that takes action: email, bookings, forms, negotiation, planning | People who want an AI that does tasks, not just answers questions | Muse app on iOS and Android, muse.ai; rolling out in the US, coming to AI glasses |
Three patterns in that table are worth noticing. First, the two model families point in opposite directions: Llama is for people who want to run a model themselves, Muse is for people who want to call Meta’s models through Meta’s own services. A developer choosing between them is choosing between self-hosting and an API, not between two versions of the same thing. Second, only one row in the table is a product a regular person opens: the assistant. The other three rows are infrastructure behind it or beside it, and confusing any of them with the assistant is how the stale “Llama powers Meta AI” claim survives. Third, the paid layer — Meta One — does not appear as a row at all, because it is not a separate product. It is a subscription that buys more of the assistant and more features in the apps, and it is covered in the pricing section near the end of this guide.
The rest of this guide follows the table in order. It explains Llama 4 and its licence, then the Muse shift and the new model family, then the assistant and the agent, then each app in turn — WhatsApp, Instagram, Facebook and Messenger — then privacy, then price, and finally a decision guide for each kind of reader. Each section says which layer it is about, and where the information runs out, this guide says that too rather than filling the gap with a guess.
Llama 4, Explained: Scout vs Maverick — and the Model That Never Arrived

Llama 4 is the current generation of Meta’s open-weight models, and it is also the last generation that powered the Meta AI assistant. It was released on 5 April 2025 in two models, Scout and Maverick, and a third model, Behemoth, was previewed at the same time as still in training. Two of those three models exist. One does not. The distinction matters, because most roundups still list three Llama 4 models as though all three were available, and one of them is not.
Scout and Maverick share the same architecture and differ mainly in scale. Both are mixture-of-experts models with 17 billion active parameters — the part of the model that actually runs on any given token — and both are natively multimodal, taking text and image as input and producing text and code as output. Both were trained with a knowledge cutoff of August 2024, and both ship in pretrained and instruction-tuned versions. The differences are in the totals and in the context windows. Scout has 109 billion total parameters across 16 experts, a 10-million-token context window, and was trained on about 40 trillion tokens. Maverick has 400 billion total parameters across 128 experts, a 1-million-token context window, and was trained on about 22 trillion tokens. Scout’s enormous context window is its headline feature: it is built to hold very long documents in mind at once. Maverick’s headline is its scale: the larger total model, built for the hardest single tasks. Both support 12 languages — Arabic, English, French, German, Hindi, Indonesian, Italian, Portuguese, Spanish, Tagalog, Thai and Vietnamese — and Meta says pre-training covered around 200 languages, with developers free to fine-tune for others.
| Model | Active / total parameters | Experts | Context window | Training tokens | Knowledge cutoff | Status |
|---|---|---|---|---|---|---|
| Llama 4 Scout | 17B / 109B | 16 | 10M tokens | ~40T | Aug 2024 | Released 5 Apr 2025, open weights |
| Llama 4 Maverick | 17B / 400B | 128 | 1M tokens | ~22T | Aug 2024 | Released 5 Apr 2025, open weights |
| Llama 4 Behemoth | 288B active / ~2T total | 16 | Not published | — | — | Previewed only, “still in training” at launch; no weights, no API, no entry in Meta’s model card as of 14 Nov 2026 |
The practical details matter for anyone thinking of running these models. Scout fits on a single H100 GPU with on-the-fly int4 quantisation, which is why it is the model developers reach for when they want a long context window on modest hardware. Maverick also ships as FP8 weights, and the FP8 version fits on a single H100 DGX host. The architecture is auto-regressive, mixture-of-experts with early fusion for native multimodality — the first Llama generation that is natively multimodal rather than having vision bolted on afterwards. Training cost, as Meta’s own model card reports it, was 7.38 million H100 GPU-hours in total — 5.0 million for Scout and 2.38 million for Maverick — with 1,999 tons of location-based CO2 equivalent and zero market-based emissions, because Meta matches 100% of the electricity with clean or renewable energy. Those figures are Meta’s own, from the model card, and they are the only training figures this guide states.
On benchmarks, this guide quotes only Meta’s own numbers from the model card, for the instruction-tuned models. On MMLU Pro, Scout scores 74.3 and Maverick 80.5. On GPQA Diamond, 57.2 against 69.8. On MMMU, 69.4 against 73.4. On MathVista, 70.7 against 73.7. On ChartQA, 88.8 against 90.0. On DocVQA, both score 94.4. On LiveCodeBench, 32.8 against 43.4. The pretrained models score 79.6 and 85.5 on MMLU, and 58.2 and 62.9 on MMLU-Pro. Maverick leads on every benchmark Meta publishes, which is what you would expect from the larger model — but the gap is narrowest where Scout’s long context helps, and widest on the reasoning-heavy tests. One benchmark this guide deliberately does not quote is the launch-day leaderboard result that was widely reported at the time: the version of Maverick that topped that leaderboard was an experimental chat-optimised variant, not the released weights, and Meta’s model card itself carries no leaderboard figure for the released models. Quoting it as though it described the model you can download would be wrong. For the concepts underneath all of this — what a mixture-of-experts model is, how context windows work, and how retrieval fits in — our explainer of BERT, LLaMA, Whisper and RAG covers the foundations in plain language, and it is the background reading for this section rather than a second version of it.
Then there is Behemoth. At launch, Meta previewed Llama 4 Behemoth as a 288-billion-active-parameter model with around 2 trillion total parameters, and described it as still in training. As of 14 November 2026, that is still the whole story: Behemoth has no weights, no API, and no entry at all in Meta’s own Llama 4 model card. This guide states exactly that, and nothing stronger. Some roundups go further than that, but no Meta statement found for this guide says more, and this guide does not repeat their stronger claims. The honest description is that Behemoth was previewed, it has never shipped, and Meta’s model card does not mention it. A reader who wants Behemoth today cannot get it, and a reader who wants to know whether it will ever arrive has no official answer to quote — only the absence of one.
One more question about Llama deserves a plain answer: is there a Llama 5? Nothing on the Meta pages read for this guide announces a Llama 5, and Meta’s developer homepage, llama.com, now leads with Muse rather than Llama — its headline is “Your next build starts with Muse”, with Muse Spark on the Meta Model API and Muse Code front and centre. This guide states only that no Llama 5 has been announced. It does not claim the Llama line is ending, because Meta has published no such statement for Llama or for Behemoth that this guide’s research found. What can be said, and what the next section explains, is that Meta’s centre of gravity has moved: the homepage that once sold Llama now sells Muse, and the API that once was the obvious place to get Meta’s models now serves Muse alone.
What “Open” Actually Means: The Llama Licence in Plain Language

Llama is often described as “open source”. That description is wrong in a way that matters, and it matters most to the people who would actually use the models. Llama 4 is released under the Llama 4 Community License, a custom commercial licence from Meta — not an open-source licence in the sense the Open Source Initiative defines the term. The weights and code are free to download and use, including commercially, but the licence carries conditions, and a developer who ignores them is not licensed at all. This section states those conditions in plain language, because they are the difference between “open” as a marketing word and “open” as a legal fact.
The licence grants a non-exclusive, worldwide, non-transferable, royalty-free right to use, reproduce, distribute, copy, create derivative works of, and make modifications to the Llama materials — the weights and the code. That grant is broad. The conditions attached to it are specific, and there are five that a user needs to know.
First, the “Built with Llama” rule: if you distribute the Llama materials, a derivative of them, or a product or service containing them, you must prominently display “Built with Llama” on a related website, user interface, blog post, about page or product documentation. It is an attribution requirement that follows the model into whatever you build with it. Second, the “Llama” prefix rule: if you use Llama materials or their outputs to create, train, fine-tune or improve an AI model that you then distribute or make available, the model’s name must begin with “Llama”. You cannot take Llama’s weights, fine-tune them, and release the result under an unrelated name. Third, the attribution notice: the notice “Llama 4 is licensed under the Llama 4 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved” must be retained in a Notice text file in all distributed copies. Fourth, the 700-million-user rule: if, on the release date of 5 April 2025, the monthly active users of your products or services — or your affiliates’ — exceeded 700 million in the preceding calendar month, you must request a licence from Meta, granted at Meta’s sole discretion, and you are not authorised to use the rights until Meta expressly grants them. That clause is aimed squarely at Meta’s largest competitors, and it is the single most important difference between Llama’s licence and a genuinely open licence: the biggest companies in the world cannot simply take the weights. Fifth, the Acceptable Use Policy is incorporated into the licence by reference, use must comply with applicable laws including trade compliance, no trademark licence is granted beyond the “Llama” mark as required for compliance, and suing Meta alleging that Llama infringes your intellectual property terminates the licence. The governing law is California.
Where do you actually get the models? The licence itself names llama.com as the download route, and the models are also on Hugging Face under Meta’s meta-llama organisation — Scout and Maverick have repositories there; Behemoth does not. Meta’s developer help directs Llama users to download and self-host. That is the point that trips up new developers most often: there is no “Llama API” on Meta’s own developer platform. Cloud and hosted inference for Llama exists through third-party providers, but no hosted price was verified on a Meta page for this guide, so none is quoted here. And Meta’s own hosted API — the Meta Model API — does not serve Llama at all. Its pricing page lists only Muse models: Wonder, Spark, Ember, Image, Voice Transcribe and SAM. A developer who wants Meta’s current models through Meta’s own API gets Muse, not Llama, and the next two sections explain how that came to be and what it means in practice.
The Muse Shift: Why Meta Stopped Leading with Llama

The single most important date in this guide is 8 April 2026. On that day Meta launched Muse Spark, the first model from Meta Superintelligence Labs, and announced that it was “available today at meta.ai and the Meta AI app”. In the same announcement, Meta’s blog described Llama 4 Maverick — the model that had powered the assistant until that morning — as “our previous model”, and said Spark delivered the same capabilities “with over an order of magnitude less compute than our previous model, Llama 4 Maverick”. That sentence is the whole shift in one line: a new model family, built by a new lab inside Meta, had replaced Llama as the engine of the assistant, and it did the same work with far less computing power. Everything else in this guide follows from that day.
The shift is visible in the places a developer actually looks. Meta’s developer homepage, llama.com, now leads with Muse, not Llama: “Your next build starts with Muse. Muse Spark on Meta Model API and Muse Code.” The page lists Muse Spark 1.3, SAM 3.1, Muse Voice Transcribe, Muse Image, Muse Glimmer — noted as available on Hugging Face — and Muse Spark 1.2 and 1.1. Llama is still there, but it is no longer the headline. The Meta Model API, Meta’s own developer API, serves Muse models only — Wonder, Spark, Ember, Image, Voice Transcribe and SAM — and no Llama model appears anywhere on its pricing page. A developer arriving at Meta’s developer pages today meets Muse first, and a developer who wants Llama is directed to download the weights and self-host. That is not a small reorganisation. It is a change of strategy: Llama remains the open-weight line for people who want to run models themselves, and Muse is the line Meta actually runs, sells through its API, and puts behind its own products.
What changed for developers is straightforward to state. If you built on Llama through a hosted API, Meta’s own platform no longer offers you that route — the hosted path is Muse, on the Meta Model API, with per-token pricing that this guide sets out in the pricing section. If you built on Llama by downloading the weights, nothing changed: the weights are still published, the licence is still the Llama 4 Community License, and the download route is still llama.com and Hugging Face. If you built a product that called Meta AI itself — the assistant — what changed is the model behind it: the answers now come from Muse Spark, not from Llama 4, and the assistant’s capabilities, including its Contemplating mode, are Muse capabilities. And if you are choosing a model family today, the choice is no longer “which Llama” but “Llama or Muse” — self-hosted open weights against Meta’s hosted API — and the two have different licences, different prices, and different futures.
Why did Meta make the shift? Meta’s own announcement gives one reason in plain terms: efficiency. Spark delivers the same capabilities as Llama 4 Maverick with over an order of magnitude less compute, and it is a natively multimodal reasoning model with tool use, visual chain of thought and multi-agent orchestration — built for the agentic work that Llama 4 was not designed around. Meta also evaluated Spark under its updated Advanced AI Scaling Framework, and reported strong refusal behaviour in the biological and chemical weapons domains. A third-party evaluation by Apollo Research noted the highest rate of “evaluation awareness” Apollo had observed — the model recognising when it was being tested — and Meta said it had investigated the finding and that it was not a launch blocker. This guide reports those safety facts as Meta reported them, and it does not add a verdict of its own. The practical point for a reader is simpler: the assistant you talk to is a reasoning model with tool use, built for agents, and that design choice — not the Llama lineage — is what shapes what it can do.
Muse Spark, Wonder and Ember: The New Family in Plain Language

Muse is not one model. It is a family, and the family has three reasoning models at its centre, plus specialist models around them. The three to know are Spark, Wonder and Ember: the everyday engine, the frontier model, and the cheap one. Around them sit Muse Image, Muse Voice Transcribe, SAM 3.1, Muse Glimmer and Muse Code. This section takes them in order, with the prices Meta publishes on its developer pricing page — because for developers, the price list is the family tree.
Muse Spark is the everyday engine, and it is the model that matters most to regular users, because it is the model behind the assistant. It launched on 8 April 2026 as the first Muse-family model from Meta Superintelligence Labs: a natively multimodal reasoning model with tool use, visual chain of thought and multi-agent orchestration. It is available at meta.ai and in the Meta AI app, and through the Meta Model API in Standard and Contributor modes, in versions 1.1, 1.2 and 1.3 — with 1.3 the latest, described on llama.com as “built for long-horizon coding and agentic work”. On the API, Spark 1.3 Standard costs $1.25 per million input tokens, $0.15 for cached input, and $4.25 per million output tokens. Contributor mode costs $0.10 in, $0.002 cached, and $0.20 out — a fraction of the price, with a trade-off that the pricing section explains in full: in Contributor mode, Meta may use your prompts, code, files, context and responses to train future models. Spark’s headline capability is its Contemplating mode, which orchestrates multiple agents reasoning in parallel and is, in Meta’s words, aimed at competing with extreme reasoning modes — Meta names Gemini Deep Think and GPT Pro as the comparison. Contemplating mode is rolling out gradually in meta.ai, and Meta’s own figures for it are 58% on Humanity’s Last Exam and 38% on FrontierScience Research. Those are Meta’s numbers, from Meta’s own blog, and this guide quotes no other benchmark for Spark or Wonder — no third-party scores, no leaderboard figures — because the researcher read no Meta page that publishes them. One number this guide does not state, because no Meta page read states it, is Spark 1.3’s context window. Third parties report a figure, but the developer pricing page lists prices and rate limits, not context windows, and this guide does not repeat a number it cannot source to Meta.
Muse Wonder is the frontier model. Announced on 12 October 2026, it is, in Meta’s own words, “our most capable models to date, built on the groundwork we laid with Muse Spark”, aimed at personal and professional agentic tasks. Two capabilities define it, and Meta defines both in its own announcement. Active Memory means the model carries what matters across conversations; Active Focus means it keeps working toward a goal, recovers from errors, and checks in before taking important actions. Its expert domains are law, finance and health, alongside frontier coding. Wonder is available through the Meta Model API and through Muse Code, and Meta says it “will soon power more of Muse, rolling out first to subscribers” — meaning the Muse agent, which runs on Spark today, will run on Wonder next. On the API, Wonder costs $4.00 per million input tokens, $0.20 cached, and $20.00 per million output tokens, in Standard mode only, as a preview. There is no Contributor mode for Wonder. A stronger variant, Muse Wonder Full, was announced in the same post; its access route is not stated in the announcement, and this guide does not state one. On safety, Meta says the Wonder family is trained to avoid sharing private information without permission and to turn down requests to assist in actually harmful acts, with a full safety report described as “available soon” as of 12 October 2026. For readers who want the parallel picture from other vendors, our explainer of the Claude model family covers the other side of the frontier-model market, and our explainer of the Gemini model family covers the Gemini models Meta’s own blog names as the comparison point for Contemplating mode — but neither is a comparison this guide makes, and this article does not rank Muse against either of them.
Muse Ember is the cheap one, and it is the model that changes the economics of building on Meta’s API. It is a reasoning model on the Meta Model API, and it is the cheapest per-token of the three reasoning models. Its pricing is two-tiered by request size: for requests of up to 200,000 input tokens, it costs $0.10 per million input tokens, $0.004 cached, and $0.50 per million output tokens; above 200,000 tokens, the rates rise to $0.30, $0.012 and $1.50, and the whole request is billed at the higher tier if it crosses the line. The worked example in the pricing section shows what that means in practice: a heavy session that costs $6.00 on Wonder costs $0.15 on Ember. Ember is the model for volume work where a smaller reasoning model is enough, and it is the reason the Muse family’s price range is so wide — from ten cents to four dollars per million input tokens across the three reasoning models.
The specialist models complete the family. Muse Image is an agentic image model that, in Meta’s description, plans, self-corrects and refines — and it costs a flat $0.01 per generated image on the API, regardless of prompt length or reasoning. Muse Voice Transcribe is speech-to-text with speaker diarization and turn detection, at $0.18 per hour of audio, metered in whole seconds, with zero-data-retention at parity. SAM 3.1 detects, segments and tracks objects from a text prompt, across image and video, at $2.50 per 1,000 images and $0.20 per 1,000 video frames, with a minimum of 50 frames billed per video request. Muse Glimmer is built to enable always-on, local agents, and is listed on llama.com as available on Hugging Face. Muse Code is multi-agent coding in the terminal, built for Muse Spark, installed with a single command — curl -fsSL https://dev.meta.ai/install.sh | bash — and it is one of the two routes through which Muse Wonder is available today. Web search grounding on the API costs $2.50 per 1,000 search queries, on top of the token cost. Rate limits are per team, not per API key: Wonder, Spark and Ember share 3,000 requests per minute and 4,000,000 tokens per minute; Contributor Spark variants are limited to 150 requests and 3,000,000 tokens per minute; Muse Image is limited to 150 requests per minute; Voice Transcribe to 256 concurrent streams and 16,000 streams per hour. There is no long-context premium on Wonder or Spark, and the small amount of steering context Meta injects into every prompt is not billed.
The Meta AI Assistant: The App, the Website, and the Glasses

The Meta AI assistant is the product most people mean when they say “Meta AI”, and it is worth describing on its own terms before taking it app by app. It is Meta’s general-purpose AI assistant. It first appeared inside Meta’s apps in 2023, in the Llama 2 era; it was rebuilt on Llama 4 when the standalone app launched on 29 April 2025; and it was rebuilt again on Muse Spark on 8 April 2026. That is three engines in three years, and it is why so many explainers are out of date: each rebuild kept the name and changed what was underneath it. The assistant today runs on Muse Spark. Everything in this section describes the assistant as Meta’s own pages describe it — and where Meta’s pages describe a feature as “in testing”, this guide says so rather than presenting it as settled.
The assistant lives in six places. There is the Meta AI app for iOS and Android, launched on 29 April 2025 as a standalone app, and the website at meta.ai. There is WhatsApp, Instagram, Facebook and Messenger, where the assistant appears inside the apps people already use. And there are Ray-Ban Meta glasses, for which the Meta AI app is the companion — Meta’s earlier glasses app, Meta View, was merged into it. A conversation can start on the glasses and continue in the app or on the web, through the history tab, but not the other way round, according to Meta’s launch post. The assistant is, by design, the same assistant in all of these places: the same memory, the same history, the same account. What differs is what each app lets it do, and the next three sections take those differences in turn.
The app was designed voice-first. At launch, Meta demonstrated full-duplex voice — a conversation in which you can interrupt the assistant mid-sentence, as you would a person — in the United States, Canada, Australia and New Zealand, with a visible microphone-in-use icon and a “Ready to talk” toggle in settings. You can tell Meta AI to remember certain things about you, and it can also pick up important details based on context; its answers draw on information you chose to share on Meta products, such as your profile and your likes and engagement, and personalised responses were available in the United States and Canada at launch, with stronger personalisation if Facebook and Instagram are linked in Accounts Center. There is a Discover feed, a place to share and explore how others use AI and to remix other people’s prompts — and Meta’s launch post states the control plainly: “nothing is shared to your feed unless you choose to post it”. The assistant generates and edits images through voice or text conversation, and the web interface has presets and options for style, mood, lighting and colours. It can search the web, so it can make recommendations and go deep on a topic rather than answering only from what it was trained on. And a rich document editor was being tested in select countries at launch — generating documents with text and images and exporting them as PDF — alongside document import for analysis. Both were described as “testing” in April 2025, and their current status was not verified on the Meta pages read for this guide, so this guide describes them as announced as in testing and says nothing more about them. For readers new to AI assistants generally, our guide to how to use ChatGPT is a useful companion for the basics of working with an AI chatbot, whatever model sits behind it.
Vibes is the assistant’s video feed, and it deserves its own paragraph because it is the part of Meta AI that most resembles a social product rather than a chatbot. Launched on 25 September 2025, Vibes is a feed in the Meta AI app and on meta.ai where you can create and share short-form, AI-generated videos: create one from scratch, remix someone else’s, add visuals, music and styles, and post it to the Vibes feed, send it in a direct message, or cross-post it to Instagram and Facebook Stories and Reels. It is, in effect, a feed of videos made by AI, inside an app whose main job is conversation. For readers who want the wider video-generation picture, our comparison of AI video generators covers the tools that generate video as their main job, including the specialist models the text models in this guide cannot replace. What this guide does not state about Vibes is anything beyond the launch announcement: a standalone Vibes app and any freemium plan for Vibes appear only in third-party reporting, and the verified fact is that Vibes launched inside the Meta AI app and on meta.ai on 25 September 2025.
Muse, the Agent: The Assistant That Does Things

Muse is the fourth layer, and it is the one that changes the question. The Meta AI assistant answers; Muse does. Launched on 8 September 2026, Muse is, in Meta’s own words, “a secure, private personal AI agent that proactively helps with people’s goals and suggests ideas”. It takes action: sending an email, booking travel, opening a browser, filling out forms, negotiating on your behalf. It develops a personalised plan from a goal, coordinates time and resources, and advances the work on its own. It keeps working after you close the app, and comes back when something changes or needs your approval — before sending an email, for example, or before making a purchase. The examples Meta gives make the difference concrete: selling a car for more, lowering a bill, adjusting a training plan, turning a saved Instagram recipe reel into a grocery list, remembering friends’ dietary restrictions before sending out invitations. None of those is a question with an answer. All of them are jobs with an outcome, and Muse is built to do the job.
The privacy architecture is the part of Muse that Meta describes in the most detail, and it is worth setting out carefully, because it is also the part where the assistant and the agent differ most. Muse runs on Muse Secure VM, a dedicated cloud virtual machine for each person, holding the agent and their data. A separate agent, Sentinel, approves everything before it reaches the internet and asks permission when needed. Muse has no visibility into passwords or payment methods — credentials go into secure storage — and checkout runs through Link by Stripe, with one-time-use cards and purchase protections. People choose which apps Muse connects to and how much access each one gets, and that access can be changed or disconnected at any time. People can opt out of their interactions being used to train Meta’s AI models, and people can tell Muse to “forget” specific things. And Meta states, in the launch post, that Muse “doesn’t share a person’s conversations or the data in their VM with Meta’s ad systems”. That last claim is about Muse, the agent — not about the assistant — and this guide does not extend it to the assistant, because Meta’s post does not. A Muse Confidential VM, in which the whole virtual machine is encrypted with a key only the user holds, was promised for “later this year” in the September 2026 post; no delivery confirmation was found for this guide, so it is described here as promised, not shipped.
Muse is rolling out in the United States on iOS, Android and muse.ai, and is coming to AI glasses. It is, in Meta’s words, “free for most of what people need, with subscription plans for people who want to do more” — and that is all Meta’s launch post says about the price. No plan names and no prices for Muse subscriptions were verified on the Meta pages read for this guide, and the conflicting figures in third-party coverage are not repeated here. Do not let any article tell you what Muse costs per month: the number it quotes is not from Meta. On 5 November 2026, Muse gained real-time voice conversations and expressive avatars in the Muse app on mobile, web and Mac: tap the phone icon at the top right, choose from hundreds of voices, and use the avatar or not — voice only, alongside the conversation, or full screen — while Muse keeps working as you talk. Voice is rolling out in the United States to start, with voice on AI Glasses and Muse Charm coming soon. Meta also announced Spaces for the Muse app on 6 November 2026. The announcement itself was not read in full for this guide, so no detail about Spaces is stated here beyond the name and the date. The important distinction to carry out of this section is the simplest one: the assistant answers questions, and the agent does jobs. If you ask Muse a question, it will answer it — but the reason it exists is the work it does after the answer.
Meta AI in WhatsApp: Chats, Groups, and Private Processing

WhatsApp is where most people meet Meta AI first, and it is also the app where Meta documents the privacy controls in the most detail — on WhatsApp’s own help pages, which this guide’s research read in full. The assistant appears in WhatsApp in three ways, and the rules are different for each. There is the dedicated Meta AI chat, reached from the Meta AI circle or chat on the Chats tab. There is @Meta AI in a group chat, where the assistant only reads messages that mention it or are shared with it, and does not read the rest of the group conversation. And there is Meta AI as a group member, which any member can add: it reads and responds only to messages sent after it was added, every member sees a system message and a welcome message when it joins, admins can remove it at any time, and it cannot change group settings. In a group, Meta lists the things it can help with: quick information such as weather and scores, answering questions, planning activities, running polls, making recommendations, tracking action items, summarising decisions and threads, and helping the admins of large groups. The honest line about what it can read, from WhatsApp’s own help page, is that for Meta AI added as a group member, “Meta can read new messages sent in the group to enable AI experiences” and “may use your prompts, the messages you share, and feedback to improve AI quality”. The exception is the Private Processing variant: with Meta AI with Private Processing as a group member, Meta and WhatsApp cannot see the chat.
Private Processing is the privacy core of Meta AI in WhatsApp, and it is worth understanding properly, because it is the mechanism behind the strongest claims Meta makes about the app. It is a Trusted Execution Environment: requests are sent encrypted, and, in WhatsApp’s words, “no one, not even Meta or WhatsApp, can read or access your messages”. Messages are not stored after the request completes, or after you exit an Incognito chat. Private Processing covers message summaries, writing help and Incognito chats. Message summaries work like this: tap the “{Number} unread messages” banner, and Meta AI summarises your unread messages; only you see the summary, and it disappears when you scroll away or exit the chat. Summaries are supported in English, Indonesian, Portuguese and Spanish only, at the time of reading, and they require accepting the terms and turning on Private Processing features. Web search is turned on by default in this context, but it can be turned off at any time.
The controls in WhatsApp are the most concrete of any app in this guide, because they are the ones verified on WhatsApp’s own help pages. You can opt out of specific Private Processing features in Settings — message summaries and writing help, for example — rather than taking the whole package or nothing. A group admin, or either participant in a one-to-one chat, can turn on Restricted chat, which prevents others from using Private Processing features in that chat. Incognito chats are always optional, and messages in them are not stored. You can create or download a Private Processing activity report, which is generated on your device and only you can access it. And you can request deletion of information you previously shared with Meta AI — in the Meta AI chat, typing /reset-ai deletes Meta’s copy of your AI chats, and /reset-all-ais resets every AI chat.
Two caveats close this section. First, the availability caveat that appears on every WhatsApp help page: “Meta AI and some experiences are currently only available in limited countries and might not be available to you yet, even if others in your country have access.” Second, a limit of this guide: only the existence of Incognito chats and their non-storage are verified on WhatsApp’s help pages, so no launch date for them is stated here.
Meta AI in Instagram: DMs, Images, and the End of AI Studio

Meta AI in Instagram is the assistant in your direct messages and your search bar, and it is the app where the most has changed for creators. You can open Meta AI from the search bar in DMs, or start a chat from Messages, then Compose, then Meta AI. In a private or group chat, @Meta AI answers a question, and the question and the reply are visible to everyone in the chat. The /imagine command in a chat generates an image, and /imagine yourself — or “Imagine me” — generates images of you from photos you add. There are voice conversations with Meta AI, in English only according to the guides that cite Instagram’s help pages. And from Compose, an “AI chats” list shows Meta AI and custom AIs other people built. One honest note about this section: the researcher did not open Instagram’s help pages directly, and these details come from secondary guides that cite them and were cross-checked against several independent guides that read the same pages in October and November 2026. They are stated here as the controls Meta’s help pages describe, and the settings labels may have changed since — if a step in this section does not match what you see on your phone, the help page on your device is the authority, not this guide.
The biggest change for creators is AI Studio, and it is a change of the kind this guide exists to record: on 10 August 2026, Meta said that people “will no longer be able to create new AI characters or edit existing ones as part of AI Studio”. Existing characters and chats remain active and are being upgraded to Meta’s latest model infrastructure, and creator AIs that already exist keep answering direct messages and story replies, labelled as AI. But the door is closed for new creators. Any guide, course or tutorial that still presents AI Studio as a live way to build a chatbot is describing a product that stopped accepting new characters in August 2026. If you already have a character, it keeps working. If you do not, you cannot make one, and there is no workaround in this guide to offer you, because Meta has not published one.
On control, the headline for Instagram is the same as for the other apps: no setting removes Meta AI from Instagram entirely. What you can do is mute the Meta AI thread so it stops sending notifications, and delete the chat so it leaves your inbox. Typing /reset-ai in the chat deletes Meta’s copy of your AI chats, according to Meta’s own Privacy Center guidance as reported by the guides that read it — deleting the chat on your screen alone does not delete Meta’s copy, and the reset command is the step that does. Reels translation and dubbing, comment-summary controls and “Imagine me” likeness controls are reported in secondary coverage but were not verified for this guide, and they are not described here as settled features. The practical summary for Instagram is short: use the assistant in DMs and search, generate images with /imagine, mute the thread and reset it if you want it quiet and forgotten, and do not go looking for an off switch or a new AI Studio character, because neither exists.
Meta AI in Facebook and Messenger: Search, Summaries, and the Controls

On Facebook and Messenger, Meta AI shows up in the search bar and in the feed, and it does two jobs that surprise people who only think of it as a chatbot: it answers searches, and it summarises comments. Ask Meta AI appears in the Facebook search bar, in the Feed, under some posts, and as a Meta AI chat in Messenger. The same honesty note applies here as for Instagram: the researcher did not open Facebook’s help pages directly, and the details in this section come from secondary guides that cite them, cross-checked against more than one guide that read the same pages in October and November 2026. The step-by-step paths are stated as Meta’s help pages describe them, and the labels on your screen are the authority if they have changed.
The first of the two surprises is AI Mode in Facebook search. Instead of a list of links, AI Mode gives answers built from public Groups and Reels content. It has been reported as launched on 15 June 2026 — but that date and its scope were not verified on a Meta page read for this guide, so take the date as reported, not as confirmed. The idea behind it is straightforward enough: Facebook has an enormous amount of public discussion in Groups and short video in Reels, and AI Mode turns that material into an answer rather than pointing you at it. Whether it is available to you, and what it draws on, depends on your account and your region, and this guide makes no claim beyond the reported launch.
The second surprise is comment summaries. Meta AI writes summaries under Page posts, and they are on by default for Pages. A Page owner or admin can turn them off for the whole Page under Settings & privacy, then Settings, then Posts, and the “Allow comment summaries on your posts” switch — and can also remove the summary from a specific post, after which Meta AI will not generate a new one for that post. The Page setting does not apply to posts made in Facebook groups. In Messenger, there is a message-sharing permission: turning off “Allow message sharing” in a chat’s Information and then Message permissions stops Meta AI features that share that chat’s messages, such as summaries and editing someone else’s photo. That control is reported as still rolling out, so it may not be available on your account yet. And the training question has a partial answer on Facebook: Meta’s Privacy Center has an objection form under “AI at Meta”, and under GDPR, people in the EU and UK can object to Meta using their information for generative AI on legitimate-interest grounds. Whether an objection is honoured outside the EU and UK is not established, and this guide does not promise it. What is established, for the agent rather than the assistant, is the statement in the Muse launch post that Muse does not share conversations or virtual-machine data with Meta’s ad systems — and this guide keeps that claim where it belongs, with Muse.
Privacy, Honestly: What You Can Control — and What You Can’t

This is the section most readers came for, and it is the section this guide is most careful about, because the honest answer is unsatisfying in both directions: there is no single off switch, and there are also more real controls than the “you can’t turn it off” headlines suggest. The table below is the whole picture in one place, split between what is verified on Meta’s own pages and what is reported from Meta’s help pages by guides that read them. The WhatsApp column is verified on WhatsApp’s own help pages. The Instagram, Facebook and Messenger columns rest on Meta’s help pages as checked by secondary guides in October and November 2026, and the labels may have changed since — check the help page on your device if a step does not match.
| App | Can you remove Meta AI? | What you can actually do |
|---|---|---|
| No — the button and the search entry stay | Opt out of individual Private Processing features in Settings; Restricted chat per conversation; Incognito chats, which are not stored; /reset-ai deletes Meta’s copy of AI chats; an on-device Private Processing activity report; mute or archive the Meta AI chat | |
| No | Mute the Meta AI thread; delete the chat; /reset-ai deletes Meta’s copy of your AI chats; no full off switch; AI Studio closed to new characters since 10 Aug 2026 (verified in Meta’s newsroom) | |
| No | Turn off comment summaries on your Page; remove a summary from a single post; mute the Ask Meta AI chat; object to training use via the Privacy Center (EU/UK) | |
| Messenger | No | Turn off message-sharing permissions per chat; mute the thread; /reset-all-ais deletes Meta’s stored copies of your AI conversations |
| Meta AI app / meta.ai | Yes — uninstall it, or don’t use it | Delete chats and history in the app; nothing is shared to the Discover feed unless you choose to post it |
| Muse agent | n/a — a separate product | Opt out of training use; tell it to “forget” specific memories; disconnect services at any time; no sharing of conversations or virtual-machine data with Meta’s ad systems (verified in the 8 Sep 2026 newsroom post) |
Three things in that table deserve emphasis, because they are the three places people get misled. First, the reset commands. Deleting a chat on your phone removes it from your screen; it does not delete Meta’s copy. The /reset-ai and /reset-all-ais commands are the step that asks Meta to delete its stored copy of your AI conversations, and they are the step most “how to turn it off” guides bury or omit. Use the reset command first, then delete the chat. Second, muting is not opting out. Muting a thread stops notifications and hides the conversation; it changes nothing about how Meta uses your data. Anyone who tells you that muting the chat “turns Meta AI off” is confusing silence with removal. Third, the training objection is regional. Under GDPR, people in the EU and UK can object through Meta’s Privacy Center to their information being used for generative AI, and that right is real — which is why the wider question of how AI regulation treats training data matters, and our guide to AI regulation around the world covers the rules that create it. Outside the EU and UK, no equivalent general opt-out is established, and this guide does not promise one. For the Muse agent, the training opt-out is stated in Meta’s launch post and applies wherever the agent is available — which is one more way the agent and the assistant differ.
What about the claim that Meta uses AI chats for ads? That claim was not verified on a Meta page read for this guide, so this guide states no ads-use claim about the assistant. What is verified, and stated here in full, is the opposite direction for the agent: Meta’s 8 September 2026 post says Muse does not share a person’s conversations or the data in their virtual machine with Meta’s ad systems. The distinction matters. A claim about the assistant cannot be borrowed from a statement about the agent, and a claim about the agent cannot be used to reassure you about the assistant. If the ads question matters to you, the honest answer is that the assistant’s data use is governed by Meta’s Privacy Center and your regional law, and that the agent’s is governed by the commitments in its launch post — and the two are not the same document.
The practical summary of this section fits in one paragraph. You cannot remove Meta AI from WhatsApp, Instagram, Facebook or Messenger. You can make it quiet, by muting and archiving the chat. You can make it forget, by using /reset-ai or /reset-all-ais before you delete the thread. You can stop it reading a particular conversation, with Restricted chat in WhatsApp or message-sharing permissions in Messenger. You can stop it summarising your Page’s comments, with the Page setting on Facebook. You can object to training use if you are in the EU or UK. And you can walk away from the standalone app entirely by uninstalling it. None of those steps is an off switch, and together they are not one either. But they are real controls, they are the ones Meta documents, and they are worth using in combination rather than hunting for a single setting that does not exist.
What It Costs: Free, Meta One, and the Developer API

Meta AI costs nothing at its core, and that is a fact Meta states plainly: “The core experience across our apps and Meta AI has always been free, and that’s not changing.” What has changed is that there is now a paid layer above the free experience, and a developer API with its own price list beside it. This section covers all three — free, Meta One, and the API — and it is careful about one thing above all: Meta publishes no numeric usage limits for the free tier or for Meta One, and this guide states none. If a roundup tells you the free tier allows a specific number of messages, images or voice exchanges a day, that number does not come from any Meta page read for this guide.
Meta One launched on 15 September 2026, with more than 50 features across Instagram, Facebook, WhatsApp and Meta AI, and it is expanding to Edits, AI glasses and more over time. The structure is a set of single-app plans, a bundle, and creator and business tiers. Instagram Plus costs $3.99 a month and buys expression and customisation features on Instagram. Facebook Plus costs $3.99 a month and buys customisation on Facebook and Messenger, more Reels insights and super reactions. WhatsApp Plus costs $2.99 a month and buys themes and icons, effect stickers, pinning more chats, and storage for backups, with Focus Schedules in testing. The Core bundle costs $7.99 a month and includes all three Plus plans, plus more media generation with Meta AI — creating and editing images, and generating videos powered by Muse models — plus more Restyle on Instagram, voice effects and creative tools. The Premium bundle costs $19.99 a month and includes everything in Core, “with the most room to create content with Meta AI and across our family of apps”. Above the consumer tiers sit the creator and business plans: Essential from $14.99 a month, with presence tools, more Meta Business Agent on WhatsApp for 24/7 customer replies, a verified badge, a verified WhatsApp channel and impersonation protection pending verification; Advanced from $49.99, with story scheduling up to 30 days, links in organic posts and reels, exportable analytics, deeper insights, team access, more linked devices and broadcast credits; Expert from $149; and Max from $499, with the highest feature access and Business Agent capacity for teams at scale. Meta’s announcement adds that plans, benefits, pricing and availability may vary by region, by app and by account — so the prices above are the published list prices, not a promise about what any particular account will see.
| Plan | Price (USD/month) | What Meta says it adds |
|---|---|---|
| Instagram Plus (single) | $3.99 | Expression and customisation features on Instagram |
| Facebook Plus (single) | $3.99 | Customisation on Facebook and Messenger, more Reels insights, super reactions |
| WhatsApp Plus (single) | $2.99 | Themes and icons, effect stickers, pin more chats; storage for backups; Focus Schedules in testing |
| Core (individual bundle) | $7.99 | All three Plus plans, plus more media generation with Meta AI (create and edit images, generate videos powered by Muse models), more Restyle on Instagram, voice effects, creative tools |
| Premium (individual bundle) | $19.99 | Everything in Core, “with the most room to create content with Meta AI and across our family of apps” |
| Essential (creator/business) | from $14.99 | Presence tools, more Meta Business Agent on WhatsApp, verified badge, verified WhatsApp channel, impersonation protection (pending verification) |
| Advanced | from $49.99 | Story scheduling up to 30 days, links in organic posts and reels, exportable analytics, deeper insights, team access, more linked devices, broadcast credits |
| Expert | from $149 | Higher feature access and Business Agent capacity |
| Max | from $499 | Highest feature access and Business Agent capacity for teams at scale |
What does the money buy, in plain terms? More room to create with Meta AI — more media generation, more video, more of the creative tools — and more features in the apps themselves. What it does not buy is a different assistant: the core experience is the same, free or paid, and the paid tiers expand how much of it you can use rather than changing what it is. Whether Core or Premium is worth it depends entirely on how much you generate. A person who chats with Meta AI occasionally and rarely generates an image will notice no difference. A creator who generates images and video every day will hit the free experience’s limits — whatever they are, since Meta does not publish the number — and for that person the question is whether $7.99 or $19.99 a month is worth the extra room. The same “which tier is worth paying for” question arises with every AI subscription, and readers weighing it across vendors will find our comparison of ChatGPT Free, Plus and Pro useful as the parallel decision on the other side of the market.
The developer API is a separate price list, and it is the only place in this guide where exact numbers run to the cent — because it is the only place where Meta publishes them. The Meta Model API is pay-as-you-go, with no minimums and no upfront commitment, and all prices are in US dollars per million tokens unless noted. Muse Wonder costs $4.00 per million input tokens, $0.20 for cached input, and $20.00 per million output tokens, in Standard mode only, as a preview. Muse Spark 1.3, 1.2 and 1.1 in Standard mode cost $1.25 in, $0.15 cached, and $4.25 out — except that 1.1 is Standard only. The same Spark versions in Contributor mode cost $0.10 in, $0.002 cached, and $0.20 out. Muse Ember costs $0.10 in, $0.004 cached, and $0.50 out for requests up to 200,000 tokens, and $0.30, $0.012 and $1.50 above that. Muse Image costs $0.01 per generated image. Muse Voice Transcribe costs $0.18 per hour of audio. SAM 3.1 costs $2.50 per 1,000 images and $0.20 per 1,000 video frames. Web search grounding costs $2.50 per 1,000 search queries, on top of the token cost. The worked example Meta’s own figures support is this: for one million input tokens and one hundred thousand output tokens, Wonder costs $4.00 plus $2.00, or $6.00; Spark 1.3 Standard costs $1.25 plus $0.43, or $1.68; Spark 1.3 Contributor costs $0.10 plus $0.02, or $0.12; and Ember, at up to 200,000 tokens, costs $0.10 plus $0.05, or $0.15. The Contributor price is the cheapest by a wide margin, and the reason is the trade-off: in Contributor mode, Meta may use your prompts, code, files, context and responses to train future models. Twelve cents buys a session that costs six dollars on Wonder — and the difference is what Meta gets to keep. That is a fair trade for many developers and an unacceptable one for others, and the right choice depends on what is in your prompts.
One more money fact belongs here, in a single sentence, because it is the boundary this guide was given and not a subject it covers: Meta AI is also an ads story, and our report on how AI is helping Meta outpace Google in ads covers the ads business and the quarterly numbers. This guide is the explainer map of the models, the assistant, the agent and the apps, and it does not re-litigate ad performance here.
Which Meta AI Should You Use — and Keeping This Guide Current

The decision guide below is the whole article compressed into five lines. Each line names a reader, the layer or plan to use, and the reason — and each reason traces back to a fact established earlier in this guide, not to a feeling about the names.
A casual user who chats with Meta AI in WhatsApp or Instagram. Use the assistant as it is, free, in the app you already open. Mute the Meta AI chat if you want it quiet, use /reset-ai if you want Meta’s copy of your conversations deleted, and turn on Restricted chat in WhatsApp if you want to stop others pulling the assistant into a group. Do not pay for Meta One unless you are generating images and video often enough to feel the free experience’s limits — and since Meta publishes no numeric limit, the test is your own usage, not a number in a roundup.
A creator who makes images, video and posts. Use the Meta AI app and meta.ai for image generation and Vibes video, and consider Meta One Core at $7.99 a month if you generate often — it is the tier that buys more media generation with Meta AI. Do not plan a workflow around AI Studio: character creation closed on 10 August 2026, and an existing character is the only kind you can still use. If you run a Facebook Page, turn off comment summaries under Settings & privacy, Settings, Posts, if you do not want Meta AI summarising your comments.
A business answering customers on WhatsApp. Look at the creator and business tiers of Meta One, starting with Essential from $14.99 a month, which adds more Meta Business Agent on WhatsApp for customer replies around the clock, a verified badge and a verified WhatsApp channel. The free assistant will answer your own questions; the Business Agent is the product that answers your customers’, and it is a different purchase from a Meta One subscription for yourself.
A developer choosing a model. The choice is Llama or Muse, and it is a choice between self-hosting and an API. Download Llama 4 Scout or Maverick from llama.com or Hugging Face if you want to run the weights yourself, under the Llama 4 Community License with its “Built with Llama” rule and its 700-million-user clause. Use the Meta Model API if you want Meta’s current models: Spark 1.3 Standard at $1.25 in and $4.25 out per million tokens for everyday work, Ember at $0.10 in and $0.50 out for volume, Wonder at $4.00 in and $20.00 out for the hardest agentic work, and Contributor Spark at $0.10 in and $0.20 out if you are willing to let Meta train on your prompts. For the model-picker method itself — the framework for choosing between AI models in general, beyond Meta — our guide to how to choose an AI model sets out the decision process this article applies.
A reader who wants the wider field. Meta is one vendor among several, and the natural next read is our explainer of the OpenAI model family, which covers the other side of the market without pretending this guide has compared the two — a genuine Meta-versus-OpenAI comparison is a different article, and this one does not claim to be it.
Two things remain: the churn warning, and the maintenance of this guide. On churn: this story moved twice in 2026 already. Muse Spark replaced Llama as the assistant’s engine in April, and the Muse agent, Meta One and Muse Wonder all arrived between September and October. A guide written in April 2026 would have said the assistant runs on Llama 4, that there is no paid tier, and that there is no agent — and every one of those statements would be wrong today. Any guide that describes Meta AI as Llama-powered, presents AI Studio as a live way to build a chatbot, or quotes a specific free-tier usage cap is describing a past version of this story, and it should be read as history rather than as a guide. On maintenance: this guide was checked on 14 November 2026, and the links below are the pages it was built from. Any reader who reaches this guide months from now should re-check the prices on Meta’s developer pricing page, the lineup on llama.com, and the subscription details in Meta’s newsroom before spending money on the strength of it. A model guide is a snapshot, not a permanent truth, and the honest ones say so on the page. This one does.
Sources
Checked 14 November 2026 against Meta’s own newsroom, AI blog, model card, licence and developer pricing pages. All facts in this guide come from the sources below, read in full on that date. No third-party roundup was used as a source for any price, limit, capability or specification; where third-party pages disagreed with Meta’s own pages, Meta’s pages were followed. Two limits are stated honestly: meta.ai itself could not be opened while this guide was researched, so nothing here comes from a first-hand read of the live meta.ai interface; and the Instagram, Facebook and Messenger controls in this guide rest on Meta’s help pages as checked by secondary guides that read them in October and November 2026, with the WhatsApp controls verified on WhatsApp’s own help pages.
- Meta AI blog — Introducing Muse Spark — https://ai.meta.com/blog/introducing-muse-spark-msl/
- Meta newsroom — Introducing Muse Wonder and Muse Wonder Full — https://about.fb.com/news/2026/10/introducing-muse-wonder-models/
- Meta newsroom — Introducing Muse: The World’s First Personal AI Agent — https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/
- Meta newsroom — Introducing Meta One — https://about.fb.com/news/2026/09/introducing-meta-one-subscription-service-more-features-ai/
- Meta newsroom — Give Your Muse a Voice — https://about.fb.com/news/2026/11/introducing-muse-ai-voice-expressive-avatars/
- Meta newsroom — Introducing the Meta AI App — https://about.fb.com/news/2025/04/introducing-meta-ai-app-new-way-access-ai-assistant/
- Meta newsroom — Introducing Vibes — https://about.fb.com/news/2025/09/introducing-vibes-ai-videos/
- Llama 4 model card (Meta’s own GitHub) — https://raw.githubusercontent.com/meta-llama/llama-models/main/models/llama4/MODEL_CARD.md
- Llama 4 Community License (Meta’s own GitHub) — https://github.com/meta-llama/llama-models/blob/main/models/llama4/LICENSE
- Meta Model API — Pricing and rate limits — https://dev.meta.ai/docs/pricing-rate-limits
- WhatsApp Help — About Private Processing — https://faq.whatsapp.com/2089630958184255
- WhatsApp Help — About groups with Meta AI as a member — https://faq.whatsapp.com/1504605284095230


0 Comments