Meta Muse Wonder AI Models Put a Price Tag on Zuckerberg's AI Push

Meta starts charging for frontier AI

On October 12, Meta released Muse Wonder and Muse Wonder Full, which it calls "frontier" artificial intelligence models. The company says both perform at the level of OpenAI's ChatGPT and Anthropic's Claude.1213 The bigger change is commercial. For years Meta gave its AI away and recovered the cost through advertising. This time the best capabilities sit behind monthly fees.1115

The pricing has three layers. Inside the Muse personal agent app, Muse Wonder comes with subscriptions of $20 or $100 a month. The app itself stays free, and paying users get more usage and the newest models.1514 Developers can reach Muse Wonder through the Meta Model API. The more powerful Muse Wonder Full is available only to subscribers of Muse Code, Meta's software engineering assistant, which costs $50 or $200 a month.1518

Nearly every outlet covering the launch reads it the same way: Meta is trying to build AI revenue that does not depend on ads.151617 That reading is correct. To see why the timing matters, look at Meta's income statement.

What Meta says Wonder can do

Internally, the model was codenamed "Watermelon." Chief AI Officer Alexandr Wang says it was trained to help personal and workplace agents get past real-world obstacles. If a restaurant has no tables, for example, the agent should look for other options rather than stop and ask the user what to do.1511 Meta also highlights two features. "Active Memory" keeps information about a user across conversations. "Active Focus" is meant to keep an agent working toward a goal when its first approach fails.18

Meta's performance claims are bold. It says Wonder beats Claude Opus 5.5, Claude Fable 5.1 and GPT-6 Astra on some benchmarks for coding, finance and legal work.17 The independent scores from Artificial Analysis give a more measured picture. On that firm's Intelligence Index, Wonder Full scores 56. That ties Anthropic's Claude Sonnet 5.5 and beats GPT-6 Astra at 53, but trails Claude Opus 5.5 at 58. The standard Wonder scores 52.20 Wonder Full does lead on legal tasks and long-running coding work. It falls behind on Terminal-Bench and HealthBench Pro.20

In short, Meta now has a model close to the top tier, though not at the very top. That is still a big improvement for a company whose earlier models were poorly received in Silicon Valley last year.14 It is also about 14 months after Zuckerberg reorganized Meta's AI work into Meta Superintelligence Labs, following a $14 billion investment in Scale AI that brought Wang to Meta.1620

The pricing claim doesn't fully hold up

The reporting disagrees most on pricing, and this matters a lot for anyone weighing AI software licensing costs.

Wang has said Wonder's API price will be "meaningfully below" comparable models from OpenAI and Anthropic.1518 VentureBeat's published rate card shows something narrower. Wonder costs $4 per million input tokens and $20 per million output tokens, which is exactly what Anthropic charges for Claude Opus 5.5. It is cheaper only than GPT-6 Astra and Claude Fable 5.1, which each cost $10 and $50.20 Artificial Analysis's own figures also list Wonder among the more expensive models on the market.1319

Cost per finished task complicates things further. Artificial Analysis estimates an average Intelligence Index task costs $4.52 on Wonder, $5.98 on Opus 5.5 and $3.26 on GPT-6 Astra.20 Even though GPT-6 Astra has a much higher token price, it can cost less per job because of how many tokens each model uses. Wang's "meaningfully below" claim depends on which rival and which measure you pick.

My reading: Meta is not competing on price this time. VentureBeat describes Wonder as a premium frontier product priced alongside Opus, not a budget option like the earlier Muse Spark.20 For enterprise buyers, the real attractions are Meta's distribution and its integrations, not lower licensing costs.

There is also a catch for businesses. Wonder Full, the strongest model, has no pay-as-you-go API price. Companies can try it inside Muse Code but cannot yet build it into their own products through the standard API.20 Neither model comes with downloadable weights. That breaks with the Llama approach that made Meta popular with companies running AI on their own servers.20 Customers who chose Meta because they could self-host now face the same hosted, metered setup that OpenAI and Anthropic use.

Why the earnings pressure makes this urgent

The subscription push follows a difficult few quarters for Meta's finances.

In the second quarter of 2026, revenue rose about 28% to $60.80 billion. But diluted earnings per share of $6.18 came in roughly 14% below consensus, ending a streak of six straight beats. Operating margin dropped from 43% to 31%.7 Capital spending rose 82% year over year, and free cash flow shrank to $784 million.79 One analysis notes that free cash flow was $8.55 billion a year earlier. Operating cash flow reached a record $31.9 billion, and capex of about $31.1 billion used up almost all of it.4

Advertising still pays for everything. Ads brought in $59.36 billion of the $60.80 billion in second-quarter revenue.10 Outlets report Meta's 2026 capex guidance slightly differently. Investor-relations summaries give $130 billion to $145 billion after the July update.107 Other reports use a $125 billion to $145 billion range or a $135 billion floor.31 Whichever figure is right, the amount is roughly double 2025's $72.22 billion.10

The comparison with peers is unfavorable for Meta. One analysis points out that Alphabet's cloud business gives it paying customers whose spending shows a return on its infrastructure. Meta's plan to sell AI to outside customers is still unproven.5 Subscriptions are Meta's attempt to close that gap.

Can subscriptions move the numbers?

The Muse app has had a strong start. It launched September 8 and reached 5 million downloads in 22 days, according to Sensor Tower. ChatGPT took 56 days to reach that mark, Grok 103 and Claude 492.1418 Meta also has partnerships with OpenTable, Walmart, Best Buy, Spotify and Ticketmaster, plus connectors to Canva, Asana, Zoom and Slack.1814

Downloads do not mean people are paying, though. Analysts vary widely on how much revenue this could bring. Oppenheimer's Jason Helfstein estimated Meta would need about 115 million paying Muse subscribers at $20 a month to reach roughly $28 billion in annual AI agent revenue, and he doubts that will happen.7 TD Cowen is more optimistic. It models Muse revenue growing from $55 million in 2026 to $1.8 billion in 2027 and $27 billion by 2031.9 Morgan Stanley's bullish case adds subscriptions to agentic ad tools and a possible "neocloud" computing business as drivers that could lift earnings per share by more than $10.8

The gap between these forecasts shows the core uncertainty. Meta has reach, with about 3.6 billion daily active users.16 Very small conversion rates could still produce meaningful revenue on that base. But even at its most optimistic, subscription revenue is small next to $130 billion or more in annual capex. Tens of millions of paying users would barely cover one quarter's spending.

Investors liked the launch anyway. Shares rose about 4% to $748.34 on the day.17 The next test is Meta's earnings report, expected around October 28. Consensus calls for EPS of $6.28, down from $7.25 a year earlier, on revenue of about $63.3 billion.6 One analyst also expects Meta to raise its capex outlook by another $5 billion to $10 billion.6

Trust is another hurdle

Charging for AI agents brings new risks. Some Muse users have complained on social media that the agent accessed their personal messages without permission. Meta denies this, though Wang says the incidents are being investigated.11 Meta's strongest protections against cyber misuse come only with the paid model, which means free users get weaker safeguards.15 Meta also has not said whether agents running on Wonder misbehaved during testing.18 For a product that sends emails and makes purchases for users, those gaps could slow subscriber growth.

Bottom line

Meta now has a competitive model and a clear pricing plan. Its claims of being cheaper than rivals are overstated, and its subscription goals are small compared with its spending. The launch matters mostly as a test of whether people who already use Meta's apps every day will pay for an AI agent. The October earnings report will show whether investors care more about that test or about the cost of the infrastructure behind it.