AI costs spike as subscriptions hit pricing wall — firms turn towards Chinese LLMs, open-source models to extend budget

red arrow showing downward trend — (Image credit: Getty Images)

The cost of serving AI via a subscription model has steadily increased for AI firms, especially as the decrease in cost per token has not kept pace with the spike in token usage. According to SemiAnalysis, the subscriptions that both Anthropic and OpenAI offer are much cheaper than the actual cost you have to pay if you maximize their usage. The research firm purchased every subscription from the two AI providers and discovered that the approximate maximum possible spend (assuming API pricing) is far larger than what users pay every month. For example, Claude Max 20x costs $200 a month, but maximizing it would cost $8,000 a month in token spend, while ChatGPT Pro 20x, which is also $200 monthly, has a maximum possible spend of around $14,000.

Go deeper with TH Premium: AI shortages

Anthropic breaks even on its two lower plans (Claude Pro and Claude Max 5x) at 20% utilization, while OpenAI starts losing money if utilization on its base plans (ChatGPT Plus and ChatGPT Pro 5x) exceed 11.4%. Things are much worse for the two companies’ top-end offerings, with Anthropic hitting 0% gross margin if utilization reaches 10%, while OpenAI is in the red if usage exceeds 5.7%. This is certainly unsustainable, but cutting features or raising subscription prices is likely off the table for these companies as well.

It’s not all bad news, though — as new models arrive and more data centers go online, the cost of serving existing models is bound to decrease, with SemiAnalysis predicting that serving Opus 4.8-level models at $20 a month could become profitable soon. On the other hand, frontier models, like Mythos, will still be much more expensive to run, so it’s likely that the latest, most advanced features could be reserved for API access only, meaning you’ll need to pay for it on a per-token basis.

Expensive frontier models have firms looking elsewhere

As SemiAnalysis showed, subscription tiers are more affordable than API access. However, you’d still need the latter if you want to access the full capabilities of these AI models, and this is where budgets start breaking. Powerful agentic AI uses up to a thousand times more tokens than the average model, and big firms like Microsoft, Meta, and Amazon are backing off “tokenmaxxing” as costs spiral out of control. One unnamed company even blew through $500 million in one month after failing to impose a usage limit on its employee licenses.

Because of this, some firms have started using tools that switch these expensive frontier models for cheaper, more affordable ones, including Chinese open-source models like DeepSeek. A Wall Street Journal report says costs could be reduced by up to 95% by allowing agents to switch between AI models as needed. “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.”

Flo Crivello, the founder of Lindy, a startup providing AI executive assistant services, also told WSJ that the company has moved towards DeepSeek V4, as it proved to be as capable as Sonnet while costing ten times less. Although it still reserves Anthropic’s models for advanced work like coding, Crivello said that using the cheaper model has “saved the company millions of dollars.”

Other firms have begun building their own AI using open-source models, which are tailored to their specific needs and trained on in-house data. While this might seem complicated and expensive at first, it could save the company in the long run, as it would not have to rely on third-party providers for its AI needs. Some even claim it could outperform frontier models, as they’re built for the firm's specific needs and applications.

The availability of cheaper models and AI agents that optimize operational costs by using the more expensive options only as needed is putting pressure on OpenAI and Anthropic to lower their prices. OpenAI CEO Sam Altman has talked about the issue of ballooning AI token costs and said the company is looking for ways to help users “get more value for less spend” when using ChatGPT.

Follow Tom's Hardware on Google News, or add us as a preferred source, to get our latest news, analysis, & reviews in your feeds.

TOPICS

Jowi Morales is a tech enthusiast with years of experience working in the industry. He’s been writing with several tech publications since 2021, where he’s been interested in tech hardware and consumer electronics.

4 Comments Comment from the forums

SSGBryan

These aren't going to become profitable; the capex expenditures alone assures that.
Reply
Elfcat

SSGBryan said:
These aren't going to become profitable; the capex expenditures alone assures that.
It's interesting isn't it. These AI agents are supposed to be great replacements for humans, but whaddaya know, the ones who know literally everything are too expensive for most jobs and it's really less of a headache to just have the one with the skill set your concern really needs!
Reply
Dementoss

Possibly another sign that the bursting of the AI bubble, this way comes...
Reply
tanon

It's interesting, I started developing a web-based WYSIWYG for lvgl on MicroPython this week, using GitHub Copilot (with Claude Sonnet 4.5)

I had been used to using pretty much unlimited tokens with my Copilot Annual subscription, but the recent plan changes to API based pricing meant that I ran out of my $10 premium request credits in the first 4-5 hours or so and had to add a further $15 credit to get to the stage where I had the WASM runtime and the code to get the LVGL canvas to display using SDL and Emscripten. This also involved a huge amount of back and forth with the agent, trying to fix a bug with setting up the LVGL canvas - a lot of which felt like pure guesswork from the Agent, until I was able to dig into the web console myself and point out exactly where the exceptions were occuring.

So I decided to search for cheaper options and found that Deepseek V4, whilst not natively supported by Copilot, actually had a VSCode Extension, which essentially gave it native support. I loaded up $20 of credit with Deepseek's API platform and 20 hours of work and 305M tokens later, I have only spent around $3.50 and implemented probably 4-5x the amount of work as I did with Sonnet.

What's even more impressive is that most of the code created by the Deepseek agent actually just worked first time, unlike the stuff implemented by Sonnet, which almost universally required 3-4 further prompts to refine or fix bugs.

I think this is because the Deepseek architecture actually has the agent go back and check whether the first output is correct or not, or whether it has made mistakes in its "reasoning". You see this all the time in the chat prompt, where it will say "Wait, actually", or "But the user said..." etc.

Another very cool thing was that with the Deepseek agent, I was actually able to share the VSCode Browser window directly with the agent, allowing it to directly manipulate it with simulated key strokes and mouse clicks using Playwright, allowing it to do most of the debugging itself, directly. This feature never showed up when I was using Sonnet, so I don't know if that was something that was added with the extension, or just a recent update to VSCode?

Anyhow, based on the current pricing, I doubt I will bother using any of the models included in the Copilot subscription much anymore, as they are way too expensive to actually get any meaningful amount of work done and don't seem to produce output that is significantly better (if at all) than that of cheaper competitors.
Reply