Is the Era of Affordable AI Over?

In Brief

New analysis from SemiAnalysis found that a $200 a month Claude plan can deliver roughly $8,000 in compute and a $200 ChatGPT plan up to $14,000, meaning the AI labs are heavily subsidizing their power users. With Meta now curbing employee AI usage and OpenAI weighing drastic price cuts, the open question for every business is whether these plans are about to get much more expensive or much cheaper.

What Happened

The era of cheap AI may not be what it seems, and the answer cuts both ways. The analyst firm SemiAnalysis bought one of each top Anthropic and OpenAI subscription, then ran long coding tasks until they hit the weekly limits. The finding: a $200 a month Claude plan delivered up to about $8,000 worth of tokens at API pricing, and a $200 ChatGPT plan reached roughly $14,000. Tokens are the unit AI companies bill by, every word in and every word out. API pricing is what you would pay per use without a flat monthly plan. In other words, the labs are heavily subsidizing the people who use these plans the hardest.

That subsidy is starting to bite. The Information reported that Meta is moving to curb employee AI usage as its internal AI costs climb into the billions. A memo went to roughly 6,000 employees flagging an "exponential increase" in usage, and starting in 2027 Meta plans token budgets, a dashboard called AI Gateway, and a push toward its own MetaCode assistant. At the same time, The Wall Street Journal reported that OpenAI is considering, though not committing to, drastic price cuts as it anticipates a war for users with Anthropic. So which way does this go, a 10x spike or a price war to the floor? SmarterX founder and CEO Paul Roetzer broke it down on Episode 219 of The Artificial Intelligence Show.

The Key Numbers

$8,000 - Monthly compute value of a $200 Claude plan at API pricing

$14,000 - Monthly compute value of a $200 ChatGPT plan at API pricing

$200 - Monthly price of the plans SemiAnalysis tested

-900% - Return to the lab when a Max user maxes out usage

$20 - Monthly price SemiAnalysis says could soon profitably serve an Opus 4.8-level model

Why No One Knows How to Price This Yet

The cost swings wildly with usage. The chart behind the SemiAnalysis post tells a powerful story. A Max-plan user touching only 5% of their allotment is pure profit, with huge margins for the lab. Usage around 10% breaks even. Push it to 100% and the lab is staring at a negative 900% return on that customer. So the same plan can be a moneymaker or a money pit depending entirely on who is holding it, which is why pricing these products is so fraught.

"It kind of shows why there's just so much uncertainty around how to price these things," says Roetzer.

The light users may be funding everyone else. The analysis shows what power users consume, but it does not show how few people come anywhere near those limits. "Let's say the top 5 or 10% of users are probably using 90% plus of all the compute or all the tokens in a given week or month," says Roetzer. "So there's probably a whole bunch of money being made by these labs for people who aren't using anywhere near that level of tokens." The heavy users get subsidized, in other words, but the bill may be quietly covered by the person paying twenty bucks a month who opens the app five times.

Google holds the pricing trump card. When every other lab is raising money to underwrite this, one player is not.

"I keep coming back to Google's advantage in all of this," says Roetzer. "If anybody has price flexibility, it's Google, because it's a wildly profitable and financially secure company, while everyone else is trying to raise money to underwrite all of this." If a price war breaks out, Google can sustain lower prices longer than rivals burning investor cash to stay in the game.

The floor keeps dropping. The reason this might resolve toward cheaper, not pricier, is efficiency. SemiAnalysis itself wrote that "the rapidly falling cost of intelligence means you'll be able to profitably serve something like Opus 4.8 level models for $20 a month in the near future." That makes a wholesale rug-pull on flat plans less likely, because serving these models is getting cheaper, not more expensive. The catch SemiAnalysis flags is that labs may simply withhold the newest features and models from subscription tiers instead.

"By this fall, when we're all in budgeting for 2027, how do you even begin to do this? It's going to be crazy."

— Paul Roetzer, founder and CEO of SmarterX, Episode 219 of The Artificial Intelligence Show

SmarterX Take

The hard part for business leaders is that the variable everyone budgets against is moving faster than the budget cycle. Most teams will set their 2027 AI spend this fall, and they will be guessing at the price of a capability that could be 10x cheaper, sold under a totally different model, or partially fenced off behind premium tiers by the time the year starts. The uncomfortable truth underneath the math is that a small fraction of staff drives the vast majority of consumption, so any per-token budget built on an average across the whole company will be wrong for almost everyone in it. The CIOs and CFOs trying to forecast this are in for a world of hurt.

What to Watch

OpenAI's pricing decision is the domino everyone is staring at. If OpenAI does cut prices to win users from Anthropic, it could trigger a broader price war that pulls flat-plan costs down across the board. The flip side is the SemiAnalysis warning that labs may withhold their best new models from subscriptions to protect margins, which would leave the headline price flat while the value quietly erodes.

Open-source models could reset the whole calculation. If an Opus 4.8-level open-source model arrives within six months, the cost of serving frontier-grade intelligence collapses and the subsidy problem largely solves itself. That is the single biggest wildcard heading into 2027 budgeting season, and it is the reason locking in today's assumptions is a trap. Track how fast open models close the gap, because that pace, more than any lab's pricing page, determines where this lands.

Which AI Platforms Companies Are Actually Paying For

Pricing wars only matter to the extent that organizations are buying, and the buying is splitting sharply by company size. According to the 2026 State of AI for Business Report, 59% of organizations provide ChatGPT, making it the most widely licensed platform overall. But the picture flips by size: 73% of small firms use ChatGPT, while the same share of large enterprises standardize on Microsoft Copilot. Claude (37%) and Gemini (42%) are meaningful but still trail the leaders.

That split is exactly why the subsidy question lands differently depending on where you sit. Smaller firms picking tools by leader preference feel price changes immediately, while enterprises buried in Office and Azure relationships absorb them through existing contracts. The report is built on more than 2,100 responses across roles, functions, and industries, and it is the clearest benchmark for understanding which platforms companies are actually paying for and why. Read the full report →