Three assets Meta actually owns
Strip away the product announcements and Meta's position rests on advertising cash flow, distribution across Instagram, Facebook and WhatsApp, and the data those platforms generate. None of the three can be bought quickly by a pure-play lab, because none of them are technology problems.
The spending that follows from that position is large. Mark Zuckerberg has said Meta was prepared to commit between $100 billion and $200 billion to AI development. Part of that is already contracted: a compute deal with Nebius Group worth $27 billion, split into $12 billion of dedicated capacity and a $15 billion expansion.
Meta is also building its own silicon. The MTIA chips developed with Broadcom are estimated to save around $8.5 billion in infrastructure costs. That is a different kind of number from an API price cut. It lowers the cost of running models Meta already owns.
The other side of the ledger
Anthropic carries an estimated $518 billion in compute and cloud obligations. OpenAI and Anthropic both depend on subscription and API revenue to service that kind of commitment, which means their price lists are also their business models.
Meta's are not. If Meta cuts the price of an API tomorrow, advertising still pays the electricity bill. If a pure-play lab cuts prices, it cuts the thing that funds the compute.
That asymmetry shows up in how the two sides talk about their products. Meta pushes agentic systems aimed at commerce and business workflows, where its existing platform data is an input. OpenAI, Anthropic, Google and Mistral push benchmark scores and token economics, because that is what their buyers compare.
What the token prices look like now
These are the list rates for the current frontier generation, in USD per million tokens. European customers pay local VAT on top, and the EUR figure moves with the exchange rate, so the dollar list is the only stable comparison point.
| Model | Input / 1M | Output / 1M |
|---|---|---|
| GPT-6.1 Sol (OpenAI) | $2.00 | $10.00 |
| Claude Haiku 5.5 (Anthropic), up to 100K | $0.10 | $0.50 |
| Gemini 4 Argon (Google), intro rate | $2.00 | $10.00 |
| Grok 4.7 (xAI), under 200K | $2.00 | $6.00 |
| Mistral Large 4 (intro / standard) | $0.68 / $1.36 | $2.09 / $4.18 |
| DeepSeek-V4.1-Flash (peak / off-peak) | $0.30 / $0.15 | $1.20 / $0.60 |
| GLM-5.3-Flash (Zhipu AI) | $0.15 | $0.50 |
| Muse Wonder (Meta) | not published | not published |
Meta's Muse Wonder, released on 12 October 2026, has no publicly detailed per-token price for the hosted series. That is the gap a developer notices first. You can budget a build on GPT, Claude, Gemini, Mistral, DeepSeek or GLM from a pricing page. For Meta's hosted models you currently cannot.
What changed in the last three months
Meta put out the Muse model line and started selling a multi-platform subscription, Meta One, across Instagram, Facebook and WhatsApp. That service recorded 15 million initial signups. It also expanded its compute base through large infrastructure contracts and team acquisitions to support agentic workloads.
On the other side, OpenAI released GPT-6.1 Sol on 22 September, Google made Gemini 4 Argon generally available on 13 October, and Mistral pushed Mistral Large 4 into public preview on 6 October at introductory rates well below the US labs. The European option remains the cheapest Western model at the top of that table.
The EU rules apply to Meta too
General-purpose AI in the European market is no longer governed by voluntary guidance. Providers now face binding statutory requirements, systemic-risk controls and Article 50 transparency obligations under the EU AI Act. Any Meta model offered to European users falls under the same regime, which means compliance work sits on top of the model work, regardless of how the company funds it.
For buyers, the practical question is not who has more money. It is which vendor's business survives a bad quarter. A lab funded by API revenue has to defend its price list. A lab funded by ads does not.
We keep a running comparison of hosted API rates against local model costs in AI Arena, and the sub-dollar input tier is where most production decisions actually get made. Meta has not entered that tier with a published number yet. More context on the model releases behind this table is in our magazine archive.
Meta's last published infrastructure commitment was the Nebius contract at $27 billion. Muse Wonder's hosted price remains unlisted.
Can European developers build on Meta's Muse models today?
Meta has not published per-token pricing for the hosted Muse series, so cost planning is not possible from a public rate card. Check per-country terms before committing a production build.
Is Meta One available in EU countries?
Meta One spans Instagram, Facebook and WhatsApp and recorded 15 million initial signups. Subscription pricing and rollout vary by market, so the country-level terms are what matter for European users.
Does Meta's advertising funding model make its AI cheaper?
Not automatically. It makes Meta less dependent on API revenue to cover compute. Whether that becomes a lower price depends on what Meta publishes, and so far it has published nothing per token.
