Meta shipped a frontier model. The catch is which one you actually get.

You've probably already seen Meta AI. It's the little blue circle that started showing up in your WhatsApp chat window a while back. The suggestions in Instagram's search. The assistant embedded in Facebook Messenger, in Ray-Ban smart glasses, in Meta's website if you ever visit it.

That Meta AI just got a way better brain. Two of them, actually. And which one you get depends on how much you pay.

On October 12, Meta launched Muse Wonder and Muse Wonder Full — the first models from Meta Superintelligence Labs, the unit Meta formed this year after pulling in researchers from the teams who built GPT-4 and Gemini. The models are genuinely frontier-class, meaning they're competitive with the best models OpenAI and Anthropic have released. The catch is how Meta is distributing them.

The split, explained

There are two models and they behave like two different products.

Muse Wonder is the one you can access via API. $4 per million input tokens, $20 per million output tokens. One million token context window. If you build AI products, this is the model you can actually integrate today.

Muse Wonder Full is the better model. According to Artificial Analysis, it scores 56 on the AA Intelligence Index v4.3, compared to 52 for Wonder and 58 for Claude Opus 5.5. Two points behind the current top model. That's genuinely close. But Wonder Full is not available via API. To get it, you need a Muse Code subscription: Power tier at $50/month or Max tier at $200/month.

So: the flagship model that competes with the top of the industry is locked behind a subscription product. The API gets the runner-up.

What does this mean for the Meta AI you already have?

Okay so this is the part that matters for most people reading this.

Meta AI in WhatsApp, Instagram, and Messenger is getting an upgrade. The new Muse models power the updated Meta AI assistant across all those platforms. If you're a regular Meta AI user, you'll notice better answers on complex questions, better memory across conversations (more on that in a second), and better handling of multi-step tasks.

You don't need a subscription for that. The apps you already have just get better.

The subscription is for Muse Code, which is Meta's developer-and-power-user product: a coding assistant that runs on Wonder Full, plus API access for people building things. Similar positioning to GitHub Copilot or Claude.ai. A different audience than WhatsApp users.

Active Memory and Active Focus

Two new features shipped alongside the models and they're worth understanding separately from the benchmark numbers.

Active Memory means the model can retain context across separate conversations. Not just within a chat session — across sessions over time. If you told Meta AI last week that you're allergic to shellfish, that information persists into this week's conversation about restaurants. This is a real change in how AI assistants work, and it's the feature that makes the everyday-app use case substantially more useful.

Worth noting: this raises the obvious privacy question. Meta's systems retaining your preferences and conversation history across time is exactly what makes it useful, and also exactly what gives people pause. The data stays in Meta's systems. That's the tradeoff.

Active Focus is the agent-mode feature. It lets Wonder Full run multi-step tasks without constant check-ins — research a topic across multiple sources, draft a document, check its own work, return a finished result. This is aimed at the same workflows that Claude Projects and ChatGPT's Tasks are targeting.

Source spread

• TechCrunch — hype. Strong on the frontier positioning and Meta Superintelligence Labs context; quotes Meta on compute investments.
• VentureBeat — skeptic. Leads with the Wonder/Wonder Full split and the API limitation; notes the benchmark provenance question.
• Yahoo Finance — builder. Business angle on Muse Code pricing and competitive positioning vs GitHub Copilot.
• Artificial Analysis — builder. The only independent benchmark source in this story. Their AA Intelligence Index v4.3 numbers are the most reliable third-party comparison available at launch.
• Latent Space — skeptic. Technical read on Active Focus architecture and the missing evals.
• AIWeekly — builder. Pricing breakdown and the Muse Code subscription tiers.

What deserves a side-eye

• The benchmark provenance. Every number Meta published except the Artificial Analysis score comes from Meta's own evaluation suite. "Our model scores X on our benchmark" is a claim, not a result. Until independent researchers can reproduce the numbers on standard public benchmarks, the scores deserve a mental asterisk.
• The timing of the API split. Releasing the weaker model via API and keeping the flagship locked to a subscription is a deliberate strategic choice. It's not wrong, but it means the developer community is working with Wonder while the consumer experience gets Wonder Full. That gap tends to close as companies feel competitive pressure, but right now it's real.
• Active Memory's data model isn't fully documented. What gets remembered. How long it persists. How you delete it. What happens if you switch regions. Meta has published some of this, but not all of it. Worth reading the privacy settings before using the memory features heavily.

What to do about it

For everyday users:

• You don't need to do anything. If you use Meta AI in WhatsApp or Instagram, it gets better automatically. You're already on the improved model.
• Try the memory features deliberately. Active Memory is opt-in to start. Meta AI's settings let you turn it on. If you want it to remember context between conversations — your preferences, ongoing projects, relevant personal details — enable it. If handing Meta that data feels like too much, leave it off. Both are reasonable calls.
• Muse Code is for heavy users and developers, not casual users. If you're just chatting with Meta AI, the free tier is still the free tier. The $50–$200 subscription is aimed at people who want a coding assistant or API access.

Further reading

• TechCrunch — "Meta releases its first frontier model" — launch context and Meta Superintelligence Labs background
• VentureBeat — Wonder vs Wonder Full API split — the clearest summary of the tier difference
• Artificial Analysis — AA Intelligence Index v4.3 — independent benchmark data; the only third-party numbers in this story
• Latent Space — AINews: Meta Muse Wonder and Wonder Full — technical deep-dive on Active Focus architecture
• Yahoo Finance — Muse Code pricing and competitive context — business angle and positioning vs Copilot and Claude.ai