Meta priced AI at $7.99. The math loses on heavy users.

Meta One undercuts ChatGPT at $7.99 a month, but these plans break even only below 11-20% usage. Read what that means before you build on them.

The Editors · 6 min read · Jun 22, 2026

Black samsung galaxy smartphone on white surface

Meta now charges for its AI. The cheapest plan, Meta One Plus, costs $7.99 a month. The step up, Meta One Premium, costs $19.99, the same sticker as ChatGPT Plus and Google AI Pro (CNBC, TechCrunch). The $7.99 tier undercuts both of those by more than half.

Here is the part the launch coverage skipped. At these prices, the people who use the product most are the people the provider loses money on. The research firm SemiAnalysis modeled the major plans and found OpenAI starts losing money on ChatGPT Plus once a subscriber's usage climbs past 11.4% of the plan's headroom, and Anthropic breaks even on Claude Pro at around 20% (TechSpot). Past that line, each extra query costs the company more than the subscription brings in. Meta arriving at $7.99 does not fix that math. It tightens it.

So the read for anyone building on these tools is short: the price you see today is a floor, not a ceiling.

What Meta actually launched

Meta is testing two consumer AI plans under a new umbrella brand, Meta One. Meta One Plus is $7.99 a month and Meta One Premium is $19.99. Both start with the same features; Premium buys more capacity on heavy queries, deeper reasoning, and more image and video generation across Meta's apps (CNBC). Testing began in Singapore, Guatemala, and Bolivia.

Meta framed the move as following the same path as other model providers that charge for extra compute. The clearer reason sits in its investor messaging: the subscriptions are there to offset the hundreds of billions Meta is sinking into AI infrastructure (Euronews). This is a company that gave away a chatbot for free and is now trying to put a meter on it.

Why $20 a month can lose money on one user

A subscription is a flat fee. The cost behind it is not flat. Every prompt burns tokens, and tokens cost the provider real money on real hardware. So the economics of a $20 plan depend entirely on how much the subscriber uses it.

SemiAnalysis put numbers on the gap. On the top tiers the breakeven point is brutal: OpenAI crosses into negative gross margin at about 5.7% utilization, Anthropic at roughly 10% (TechSpot). The same analysis estimates that a $200 ChatGPT Pro plan, used to its full headroom, would run as much as $14,000 in equivalent API costs. Anthropic's $200 Claude Max plan has a comparable ceiling near $8,000.

Read that as a model, not as audited accounts. The point holds either way. These plans are priced for the light user who logs in twice a week. The power user who lives in the tool is a loss the provider eats, paid for by everyone who barely shows up. The business works only while most subscribers stay quiet.

The base business is already underwater

The discount lands on top of margins that were thin to begin with. In 2025, Anthropic ran a gross margin near 40% and OpenAI near 33% (AI After Hours). Those are software companies posting hardware-company margins. OpenAI burned about $9 billion in 2025, spending $1.70 for every dollar it earned, with burn projected to reach $57 billion by 2027.

Part of the squeeze is that inference keeps costing more than planned. Anthropic's inference bill ran 23% above its own forecast, which is what pulled its margin below target. When the cost of serving a request beats your own model of it, cutting the price the customer pays is the opposite of a fix.

"Prices are falling" is not the rescue it sounds like

The standard rebuttal is that compute gets cheaper every year, so the math fixes itself. Per-token prices really are dropping. The average cost of a million tokens fell from roughly $10 to $2.50 in a single year across major providers (Investing.com).

Usage rose faster. Agentic workflows at 2026 adoption levels consume multiples of what a 2024 budget assumed, so the bill per active user climbs even as the unit price falls. We wrote out that mechanic in the agent-loop math: cheaper tokens, far more of them, a bigger total. A 75% price cut does not save you when consumption goes up tenfold.

Why Meta can do this when a standalone lab cannot

Here is the honest distinction. $7.99 is survivable for Meta in a way it is not for a pure AI lab, and not because Meta's model is cheaper. Meta has an ad business and three billion-user apps to carry the cost. The subscription is customer acquisition, not a margin product. The real return is the data and the habit, with ads inside the assistant as the eventual payday.

OpenAI and Anthropic do not have that cushion. For them the subscription has to carry its own weight, which is why their pricing keeps drifting toward usage caps and metered add-ons. A land grab funded by ad revenue looks like a price war, but it is closer to a loss leader at the supermarket door. The cheapest sticker in the aisle is rarely the one selling at a profit. It is the same lock-in trade we covered when OpenAI moved to cut token prices: the low number on the shelf is bait for the dependency behind it.

What this means if you build on AI

If your product, funnel, or workflow sits on top of consumer AI plans, do not treat today's price as the baseline you can plan around. The structure points one way.

Budget on metered cost, not the subscription. If your usage would blow past a plan's breakeven, assume the price finds you eventually, through caps, throttling, or a new higher tier.
Expect ads inside the assistant. When the subscription cannot cover heavy use, attention becomes the backstop. Build as if the free and cheap tiers will carry ads.
Treat your data as the real fee. The $7.99 is the visible price. The behavioral data that trains the next model is the invisible one.
Avoid hard lock-in to one provider's pricing. Keep your prompts and pipeline portable, so a tier change does not break your unit economics overnight.

The sticker says $7.99. The cost of running these models has not moved to match it. Someone pays the difference, and the pricing is designed so that, for now, it is not you, until you use the thing enough to matter.