My Updated Meta AI Review: The Billion-User Assistant That Just Went Closed-Source

Hands-On Verdict

Let us cut through the press release fireworks. Meta AI just crossed a billion monthly active users, got rebuilt around a new closed-source model called Muse Spark, and is sitting on top of $125 to $145 billion in planned 2026 AI spending. That is a lot of zeros. The question I wanted to answer with this review is simpler: for someone who actually uses these tools daily, does the 2026 version of Meta AI justify the noise?

As of my May 2026 verification pass, this review covers the Meta AI that exists today - not the roadmap slide deck, not the Connect 2026 wishlist. Muse Spark launched on April 8, 2026 as the first model from Meta Superintelligence Labs, the group built around Alexandr Wang after Meta dropped roughly $14.3 billion on the Scale AI deal. Voice conversations powered by that same model started rolling out in mid-May. Every claim below comes from using the actual product, not from a demo video.

My rule holds: use Meta AI when it genuinely removes friction, not when you are chasing novelty. For the billion-plus people inside Meta’s ecosystem, there is real utility here. Whether that utility outweighs the shifting privacy landscape is a question every user should answer for themselves.

The Big Shift: Goodbye Llama, Hello Muse Spark

If you followed Meta AI from 2024 through 2025, you knew it as a Llama-powered assistant. First Llama 3, then Llama 4 Scout and Maverick with native multimodality, and a promised Llama 4 Behemoth that never shipped - the Wall Street Journal reported in May 2025 that Meta delayed Behemoth indefinitely over capability concerns. Then came the real plot twist.

In June 2025, Meta created Superintelligence Labs, invested $14.3 billion into Scale AI, and hired its 28-year-old founder Alexandr Wang as Chief AI Officer. Nine months later, on April 8, 2026, Meta dropped Muse Spark. It is natively multimodal, supports visual chain-of-thought reasoning, tool-use, and multi-agent orchestration. It also happens to be completely closed-source - a sharp departure from the Llama open-weight strategy that Meta had championed for years.

This matters because Llama was the open-source community’s poster child. Researchers fine-tuned it, startups built on it, and Meta positioned itself as the anti-OpenAI. Muse Spark is proprietary. You cannot download the weights. You cannot run it locally. You interact with it only through Meta’s interfaces. For the open-source AI crowd, this felt like the rug getting pulled.

Is Muse Spark better? In visual tasks - analyzing images, health-related photo interpretation, design-to-code - it punches above its weight. Benchmarks show a 94.8% score on design-to-code against Claude’s 77.3%. On pure coding and complex reasoning, however, both GPT-5.4 and Claude Opus 4.6 lead by wide margins. Muse Spark scores around 52 on multilingual understanding tests. It is competitive but not dominant.

Where Meta AI Actually Lives in 2026

The platform spread is broader than anything else in AI. Meta AI now reaches over a billion monthly active users across Facebook, Instagram, WhatsApp, Messenger, and its standalone app. That is not just big - it is roughly double what the assistant had at the start of 2025. For context, Meta’s family of apps hit 3.56 billion daily active users in Q1 2026, marking the first-ever quarterly decline from 3.58 billion, but still representing an enormous addressable base.

The Standalone App. Launched in April 2025, the Meta AI app graduated from “experiment” to full product in 2026. It has its own Discover feed where you can see how friends use AI, a dedicated Vibes tab for AI-generated short videos, and the full conversational assistant. If you want Meta AI without scrolling through Instagram, this is where you go.

Facebook. Meta AI sits in the search bar, in a dedicated tab, and in comments. The integration feels less novel than it did in 2024 - it is now just part of the furniture. That normalisation is probably the point. You open Facebook, you ask something, you move on.

Instagram. Direct message integration remains the primary touchpoint. Caption generation and hashtag suggestions are genuinely useful for creators. The image generation features now lean on Muse Spark’s visual reasoning, which produces sharper, more context-aware results than the Llama-era outputs.

WhatsApp. The most natural integration, partly because we already treat WhatsApp as a conversation space. Meta AI lives in your chat list like any contact. The new voice conversations feature allows natural back-and-forth - interrupt mid-sentence, switch languages, ask about something your camera is seeing. For cross-language family chats, the real-time translation is quietly one of the best features Meta ships.

Messenger. Group chat integration is practical for event planning, but the standout here is the voice conversation rollout. In mid-May 2026, Meta updated the assistant to handle real-time voice with interruption handling and topic switching. You talk to it like a person, and it responds in kind. The conversational quality is markedly better than the stilted voice modes we had in early 2025.

The Voice Conversation Upgrade (May 2026)

The update that dropped the week of May 12, 2026 deserves its own callout. Powered by Muse Spark, Meta AI now supports what Meta calls Voice Conversations - natural speech interaction where you can interrupt, switch topics mid-flow, and change languages without missing a beat. The assistant responds with generated voice, not pre-recorded snippets, so the rhythms feel more human.

In testing, I asked about cricket scores, switched to requesting a restaurant recommendation, interrupted to ask what language a street sign in a photo was in, and then asked for a translation - all in a single continuous voice session. It held up. The speech generation is not quite at the emotional expressiveness level of ChatGPT’s Advanced Voice Mode, but the interruption handling and multi-language fluidity are impressive for a product that is free and embedded across the entire Meta stack.

Vibes, Video, and Visual Creation

Vibes launched as an AI video feed inside the Meta AI app in September 2025 and was spun out toward a standalone app in February 2026. You scroll through a TikTok-style feed of AI-generated short videos, remix them, or create your own from text prompts. The quality is deliberately stylized - do not expect photorealistic Sora-level output - but for quick social media content, it works. Meta plans to eventually power Vibes with Muse Spark, which should improve the model’s visual generation fidelity.

Image generation inside chats has also evolved. Muse Spark understands spatial relationships better than Llama 4 did, so when you ask for “a golden retriever sitting on a red bicycle in front of a blue house with a white picket fence,” you are less likely to get a surrealist nightmare where the bicycle is inside the dog.

The Privacy Tension Nobody Wants to Talk About

Meta updated its privacy policy effective October 2025, and the implications for AI users are significant. The company can use public posts, photos, and interactions from its platforms to train AI models. European users under GDPR can opt out through a form buried in the Privacy Center. Users in the United States and most other regions have no such opt-out - the consent is built into the terms of service.

If you are reading this and thinking “that sounds like a lot of my data going into models I have no control over,” you are not alone. The opt-out process for EU users is functional but not easy to find. For everyone else, the reality is: if you use Facebook, Instagram, or WhatsApp, your public content is part of the training pipeline. Meta has not been shy about this. The January 2026 announcement that it would auto-consent users for AI training data sent ripples through privacy communities, but mainstream user behavior barely budged.

The deeper issue is that Meta’s entire business model runs on monetizing user data through advertising. Embedding an AI assistant that learns from every interaction into that same ecosystem creates a surveillance layer that dedicated assistants like Claude or Perplexity do not have. When Meta AI helps you plan a trip, it knows what you are interested in. When it helps you write a caption, it knows your tone. That data does not disappear - it feeds the advertising engine.

The May 2026 revelation that Meta had reassigned 7,000 employees into AI roles while laying off 8,000 others, combined with a new internal policy on training AI tools with employee data that sparked outrage within the company, suggests that even Meta’s own workforce is uneasy about where this train is headed.

Performance: Where It Shines, Where It Stumbles

Strengths. Visual analysis tasks are Muse Spark’s superpower. Upload a photo of a meal and ask for nutritional estimates, and the model delivers surprisingly detailed breakdowns. Point your camera at a circuit breaker panel and ask for identification - it works. The design-to-code pipeline, which turns UI mockups into functional code, scored 94.8% on the relevant benchmark, trouncing Claude’s 77.3%.

General conversational assistance - quick facts, definitions, translations, writing suggestions - is solid and fast. Real-time web search keeps answers fresh. The voice conversation quality has genuinely improved from an afterthought to a flagship feature.

Weaknesses. Pure code generation and complex multi-file programming tasks still belong to GPT-5.4 and Claude Opus 4.6. If you are a developer doing serious work, Meta AI is not your primary tool. Multi-turn conversations with deep context requirements still break down. The model sometimes loses track of earlier constraints after about 20-30 exchanges.

Benchmark analysis consistently places Muse Spark in the second tier for mathematical reasoning - beating Meta’s previous models but trailing the frontier leaders. The 52 score on multilingual benchmarks is middle-of-pack. Agentic tasks, where the AI needs to chain multiple tools and maintain state across steps, remain a gap area.

The Subscription Question Mark

Meta confirmed in January 2026 that it is testing premium subscription tiers for Instagram, Facebook, and WhatsApp. The core apps remain free, but advanced AI features - potentially including Vibes creation tools, extended context windows, and priority model access - are expected to move behind a paywall. Pricing has not been announced, though Meta Verified gives a rough benchmark at $15/month.

The Threads post from May 2026 suggesting £3.99 for one account, £6.99 for two, and £9.99 for three per month was not officially confirmed by Meta, but it aligns with the direction of travel. If premium AI features eventually cost something in the $5-10/month range per platform, that puts Meta AI in direct subscription competition with ChatGPT Plus ($20/month) and Claude Pro ($20/month), though Meta’s version would remain integrated across social platforms.

The Glasses Factor

Meta’s AI glasses deserve mention because they are increasingly where the technology feels most exciting. Ray-Ban Meta Gen 2, the Oakley Meta HSTN line, and the recently launched prescription-optimized Ray-Ban Meta Blayzer Optics all carry the Meta AI assistant. Features include hands-free calling and texting, live translation, image analysis through the camera, nutrition tracking by photographing your food, WhatsApp message summaries, and display recording (on the Ray-Ban Meta Display glasses).

The glasses are not a replacement for the phone-based assistant - they complement it. Ask “Hey Meta, what am I looking at?” while wearing them, and the glasses describe the scene through the camera, powered by the same Muse Spark visual reasoning that shines on the phone. The v5 software release notes from the week of May 14, 2026 added display recording and improved conversation focus in noisy environments.

My Updated Recommendation for 2026

Here is where I land after converging a year of use with the 2026 product reality.

Meta AI is the most accessible AI assistant on the planet. A billion people use it monthly, and most of them never had to download a separate app, create a new account, or pay a cent. For quick visual analysis, casual conversation, voice interaction with multilingual support, and social content creation, it is genuinely useful and getting better.

But Meta has made a strategic choice that changes the calculus. Muse Spark is closed-source. The company that once championed open-weight AI as a counterweight to OpenAI has now built a proprietary model that you can only access through their interfaces. Whether this is a pragmatic business decision to compete on capability or a longer-term lock-in play is debatable. It is probably both.

The privacy implications have not improved. If anything, the 2026 policy updates that auto-consent users for AI training - with meaningful opt-outs only for EU residents - make Meta AI a harder recommendation for anyone who cares about data sovereignty. You are not just using an AI assistant. You are feeding one with every interaction, and the terms of that exchange are dictated by a company that makes money from your data.

My advice: use Meta AI for what it does well - visual tasks, quick searches, translation, voice conversations, social content creation. Do not put sensitive information into it. Treat your Meta AI chats the way you would treat a public post, because the line between the two is thinner than Meta’s documentation suggests.

If you need an AI for research, coding, deep reasoning, or anything that touches private or proprietary information, use Claude or ChatGPT with a paid plan that offers clearer data handling commitments. Meta AI can be part of your toolkit, but it should not be the toolkit.