A $200 ChatGPT subscription could cost OpenAI $14,000 if you actually used it to its full potential

[Skye Jacobs]

Bottom line: The math behind AI subscriptions is starting to look uncomfortable. Flat monthly pricing helped fuel the rapid adoption of tools like ChatGPT and Claude, but new analysis suggests those fees may not come close to covering the actual cost of heavy use. As users push these systems harder and more demanding AI workflows take hold, the gap between revenue and compute costs is becoming difficult to ignore.

SemiAnalysis has calculated how big that gap really is. After testing subscription tiers from both OpenAI and Anthropic – running long-horizon coding and agentic tasks until weekly limits were exhausted – the firm found that the cost of theoretical maximum usage of these plans if priced at standard API rates far exceeds what users actually pay.

A $200 ChatGPT Pro 20x subscription could cost as much as $14,000 in API pricing if fully utilized. Anthropic's Claude Max 20x plan, also priced at $200 per month, has a comparable ceiling, with potential usage totaling roughly $8,000 in token costs.

Those figures help explain why utilization rates matter so much to the AI companies offering them. According to SemiAnalysis, Anthropic breaks even on Claude Pro and Claude Max 5x at around 20% utilization. OpenAI's margin is thinner. It begins losing money on ChatGPT Plus and ChatGPT Pro 5x once usage climbs above 11.4%.

– SemiAnalysis (@SemiAnalysis_) June 10, 2026

The economics get tighter at the high end. Anthropic reaches zero gross margin at roughly 10% utilization on its top-tier plans, while OpenAI crosses into negative territory at just 5.7%. It doesn't take extreme use for these subscriptions to turn unprofitable.

Adjusting pricing or restricting access is not a straightforward fix. Subscription models have been central to user growth, and pulling back risks slowing momentum in a market where capabilities remain a key competitive differentiator.

Part of the pressure comes from how AI is actually being used. Token consumption is rising quickly, especially with agentic systems that can require up to 1,000 times more tokens than a standard prompt. That kind of demand is already forcing large organizations to rethink how freely these tools should be deployed.

Microsoft, Meta, and Amazon have reportedly pulled back from internal efforts that encouraged heavy usage after costs escalated. In one widely cited example, a company burned through $500 million in a single month using Anthropic's Claude, largely because it failed to put limits on employee access.

That kind of overspending is pushing companies toward more controlled approaches. One strategy gaining traction is to shift workloads between models depending on the task. More complex queries go to expensive frontier models, while routine work is handled by cheaper alternatives.

– Dong Ming (@dming) June 14, 2026

The savings can be substantial. A Wall Street Journal report found that routing tasks this way can cut costs by up to 95%. "You don't need a model that knows quantum gravity," Columbia University vice dean Vishal Misra told the publication. "These open-source models are very capable, and the ability to charge a big premium for AI is going to diminish."

Some companies have already made the shift. Flo Crivello, founder and CEO of AI assistant startup Lindy, announced that the company moved 100% of its traffic to DeepSeek V4, switching entirely away from Anthropic's models. DeepSeek V4 proved comparable to Claude Sonnet at a fraction of the cost, and the move has "saved the company millions of dollars," Crivello said.

Others are going further by building their own AI systems on top of open-source models trained on internal data. While that requires more upfront investment, it offers tighter cost control and reduces dependence on third-party providers. In some cases, these tailored systems may even outperform general-purpose frontier models for specific use cases.

There is some expectation that costs will ease over time. As infrastructure expands and newer models replace older ones, the cost of running mid-tier systems should decline. SemiAnalysis suggests that models at the Opus 4.8 level could eventually be delivered profitably for around $20 per month.

That does not apply to the most advanced systems, though. Frontier models, including those still in development, remain expensive to run. Their highest-end capabilities may increasingly be priced via APIs rather than bundled into consumer subscriptions.

For now, AI providers are juggling two forces: users want powerful tools at low, predictable monthly prices, but the infrastructure to run them remains costly and highly sensitive to usage. OpenAI CEO Sam Altman has acknowledged the tension, noting that rising token costs are becoming a serious issue and that the company is working to help users "get more value for less spend" when using ChatGPT.

Permalink to story:

A $200 ChatGPT subscription could cost OpenAI $14,000 if you actually used it to its full potential

 

[human7]

One story is that subscriptions are too cheap, the other is that API prices are too expensive, and both are probably true. Either way, open source models are a great alternative, the best ones these days just happen to be Chinese. Not sure why American or other western firms don't put more effort into making open source models. Llama could and should have been updated several times since the lackluster release of 4 by now.

 

[ScottSoapbox]

Yeah but not everyone can keep AI pegged at 100% even during work hours. My own use case requires diligent supervision so when I'm off sleeping or motorcycling it's sitting idle. If you aren't constantly running agents to constantly prompt the AI company is doing fine.

The other factor is that inferencing is getting cheaper. More efficient models, processes, and hardware will keep driving down the cost.

Finally supply and demand. As AI companies adjust their tokens to realistic prices it incentivizes the biggest users to A use less or B switch to local models.

 

[yRaz]

One story is that subscriptions are too cheap, the other is that API prices are too expensive, and both are probably true. Either way, open source models are a great alternative, the best ones these days just happen to be Chinese. Not sure why American or other western firms don't put more effort into making open source models. Llama could and should have been updated several times since the lackluster release of 4 by now.

There are so many open and free models that when the true cost of AI decides to show it's ugly head, people who seriously use it will be better off running it locally on a $50,000 workstation. Wasn't there a story recently about a company who got hacked and ended up using something like $84,000 in Gemini tokens in a few days? And let's not forget about the privacy concerns associated with using cloud models.

I'm not anti AI, but I do think something is seriously wrong with their business models. Frankly, I see the end game of going back to a 90s software model. They train these massive models in the data center and sell you a local license to run on your own hardware. Making the model is the hard part. Once it's trained it takes a relatively small amount of hardware to just run.

 

[yRaz]

Yeah but not everyone can keep AI pegged at 100% even during work hours. My own use case requires diligent supervision so when I'm off sleeping or motorcycling it's sitting idle. If you aren't constantly running agents to constantly prompt the AI company is doing fine.

The other factor is that inferencing is getting cheaper. More efficient models, processes, and hardware will keep driving down the cost.

Finally supply and demand. As AI companies adjust their tokens to realistic prices it incentivizes the biggest users to A use less or B switch to local models.

You'd be surprised how easy it is to burn through your token usage. I've been experimenting with making AI agents and I can burn through my 5 hour limit on Gemini in just a few messages. Granted, I'm not exactly just chatting, I'm actually trying to make AI employees for a business idea I have. I'm also only paying for $20/m plan, but considering that their $200/m plan only gives me 40x more tokens for 10x the price, I'm fairly certain that this stuff could be burned through pretty easily.

This is one of those hobby project theoreticals that I'm not taking seriously and am not willing to spend more than $20/m on. I hit my 5 hour limit and let my "agents"(bots) hang out in discord and solve a problem over a couple of days/week instead of what might take them 10 minutes if I wanted to pay for the higher teir. It's a hack job, home lab project.

Anyway, before I get to far into this tangent, it's really easy to burn through tokens if you have a real application for it

 

[WhiteLeaff]

Either they have no idea what they're talking about, or they know full well and are manipulating information to create a narrative

 

[toooooot]

If true that is insane. But it is not something they cannot solve.
We already have throttling for cellphone data, AI companies will introduce something like that.

 

[Theinsanegamer]

If true that is insane. But it is not something they cannot solve.
We already have throttling for cellphone data, AI companies will introduce something like that.

And that throttling will make the tool utterly useless, much like phone data once you run out.

 

[Squid Surprise]

You don’t think they’ve already thought of this? It’s not a matter of “if” your subscription is cheaper than the cost of the service, but how many users actually surpass the value of their subscription.

It’s like going to an all-you-can-eat buffet and eating 10 pounds of shrimp… YOU might be getting your money’s worth - but how many are filling up on bread?

 

[Septimus]

One story is that subscriptions are too cheap, the other is that API prices are too expensive, and both are probably true. Either way, open source models are a great alternative, the best ones these days just happen to be Chinese. Not sure why American or other western firms don't put more effort into making open source models. Llama could and should have been updated several times since the lackluster release of 4 by now.

Take a guess....

Dependency. China puts out good cheap models, it is easy to start manipulating information on a global scale through it. Slow burn.

 

[brucek]

At current subscription pricing, I feel like I'm renting the hardware at a huge discount and getting the model for free. Makes for an easy decision for now.

The high hardware costs are a problem for the moment and I'm not sure if the same performance boost that we got over the first few decades of the personal computer era is available again at the start of the AI era. But we will get some increases and costs will decrease as more capacity comes online.

 

[Tech420]

There are so many open and free models that when the true cost of AI decides to show it's ugly head, people who seriously use it will be better off running it locally on a $50,000 workstation. Wasn't there a story recently about a company who got hacked and ended up using something like $84,000 in Gemini tokens in a few days? And let's not forget about the privacy concerns associated with using cloud models.

I'm not anti AI, but I do think something is seriously wrong with their business models. Frankly, I see the end game of going back to a 90s software model. They train these massive models in the data center and sell you a local license to run on your own hardware. Making the model is the hard part. Once it's trained it takes a relatively small amount of hardware to just run.

You would need a pretty powerful computer to run a model that has decent capabilities. This is just offsetting the water usage from data centers to consumers. Everyone will need to hire a plumber to run a water hose to their PC which will be great business for plumbers, but not very cost effective for the consumer, especially if the water rates are high in the area.

 

[pnntmp]

A tiny amount of subscribers use the "full potential" of their subscription.
What matters are the averages, not the extreme cases.

High token price is something new, once it's here the market will adapt to it quickly. There's plenty of room for optimization, it simply wasn't necessary with super-cheap fixed price subscriptions. Now it will happen in various ways. Most tasks being solved with AI don't need the cutting edge models.

By the way, do you remember the "token maxing" nonsense some tried to spin just 2-3 weeks ago?

 

[yRaz]

You would need a pretty powerful computer to run a model that has decent capabilities. This is just offsetting the water usage from data centers to consumers. Everyone will need to hire a plumber to run a water hose to their PC which will be great business for plumbers, but not very cost effective for the consumer, especially if the water rates are high in the area.

I don't know how to say this, but that's not how it works. They run cooling lines outside to massive evaporative coolers in these data centers where they spray freshwater on cooling lines and blow air over it to accelerate the evaporation. If you were a business that had maybe 1 rack filled with AI hardware instead of several football fields full of racks, that kind of cooling is unnecessary

 

[Theinsanegamer]

A tiny amount of subscribers use the "full potential" of their subscription.
What matters are the averages, not the extreme cases.

High token price is something new, once it's here the market will adapt to it quickly. There's plenty of room for optimization, it simply wasn't necessary with super-cheap fixed price subscriptions. Now it will happen in various ways. Most tasks being solved with AI don't need the cutting edge models.

By the way, do you remember the "token maxing" nonsense some tried to spin just 2-3 weeks ago?

The problem with your assertion is that you're using the buffet model, where everyone pays to get in and eat. But that isn't true in AI. Who are these mythical subscribers who pay these license costs then don't use the software?

Most people don't use AI at all. Most light users of AI don't pay for it, they use the free models. What AI doe sis closer to the YouTube model, except unlike YouTube it doesn't have the massive advertisement pool feeding it.

If what you were saying was true, then these AI companies wouldn't be blowing billions just to remain operational.

 

[pnntmp]

The problem with your assertion is that you're using the buffet model, where everyone pays to get in and eat. But that isn't true in AI. Who are these mythical subscribers who pay these license costs then don't use the software?

Most people don't use AI at all. Most light users of AI don't pay for it, they use the free models. What AI doe sis closer to the YouTube model, except unlike YouTube it doesn't have the massive advertisement pool feeding it.

If what you were saying was true, then these AI companies wouldn't be blowing billions just to remain operational.

I'm one of those subscribers, and I don't think I'm mythical in any way. I use AI a lot, but I simply use it as much as I need it, not as much as it's theoretically possible.
The "mythical subscribers" don't pay and then not use it, they pay and use it, but the price of that usage for the service provider is way below the price paid for the subscription - so the service provider makes a profit. If the provider makes a profit from 98% of the users, they can afford to operate at a loss for the remaining 2%.
The buffet model works perfectly fine in AI.

 

[NumberNine]

Everything that goes up must come down.

 

[BadThad]

This is to be as expected. I find it hard to believe that some really smart bean counters didn't already examine this thoroughly. AI is new, all these pricing issues will work themselves out over time.

 

[acrossthestars]

At some point, the bill is going to come due. Shareholders are gonna ask for their money back and these companies are not going to have nothing to show for all of the billions of dollars they've incinerated.

Making snazzier-looking gaming expo presentations might technically be a good use of AI, because it amplifies work that was already being done, but that is merely hundreds of thousands of dollars' worth of VC money. The money actually being spent on AI right now is orders of magnitude more: hundreds of BILLIONS of dollars. Nothing short of a 90% unemployment rate AND 10x ROI is going to justify these obscene spending rounds.

The math ain't mathing...

 

[noblody]

Probably most of the users use it for at most 20% of the time during working hours and even then the usage is not maxed because you need to check what it generates. With new models they are in the process of "fixing" it for the hard-core users. New models consume a lot more of the allocated quota, other models are available only at api rates. we should probably not loose too much sleep for the finances of billion dollar companies.

 

[OortCloud]

I've said it so many times - the LLM way of faking intelligence is such a barn-door engineered brute-force method and such a technological dead end. It will never lead to a true AI. That's why statements by Jensen Huang saying they have achieved AGI are such horsesh1t. I can't decide if he's an ignorant fool or just full of sh1t. Probably both.

 

[Squid Surprise]

I can't decide if he's an ignorant fool or just full of sh1t. Probably both.

Yeah, CEO of a multi-trillion dollar company he built from the ground up… must be an ignorant fool…

 

[OortCloud]

Yeah, CEO of a multi-trillion dollar company he built from the ground up… must be an ignorant fool…

So full of sh1t then?
Edit: Actually wasn't he recorded stealing money out of people's wallets and laughing about how poor they were at Computex? - so definitely ignorant too.

 

[Squid Surprise]

So full of sh1t then?
Edit: Actually wasn't he recorded stealing money out of people's wallets and laughing about how poor they were at Computex? - so definitely ignorant too.

No - he took bills out of a journalists wallet, signed them, and then handed them out - the journalist ended up with more money than he had started with along with merch… it was a publicity stunt…

 

[OortCloud]

No - he took bills out of a journalists wallet, signed them, and then handed them out - the journalist ended up with more money than he had started with along with merch… it was a publicity stunt…

It wasn't. It was a fan who handed him his wallet to sign. Instead Jensen opened it and said 'wow you don't have much money do you?', then proceeded to give his money to show girls. He gave it back later but shows a level of tone deafness I'd call ignorant.