Club Benefits
If you pay for ChatGPT or Claude and you genuinely lean on it by spending your day coding, running agents or even churning through long research sessions, then you're getting one of the best deals in tech right now. And yet, you're also the reason the deal won't last.
At least that's the uncomfortable takeaway from a new analysis by research firm SemiAnalysis, which bought every paid tier from OpenAI and Anthropic and deliberately ran them into the ground. The team pushed each plan with long, agent-style coding tasks until it hit the weekly usage cap, then worked out what all those tokens would have cost at standard pay-as-you-go API rates.
The gap between what you pay and what you get turned out to be enormous.
What the numbers actually say
The long-standing rule of thumb was that a $200-a-month plan tops out at around $2,000 of usage. SemiAnalysis found the real ceiling is far higher. In fact, in the most extreme case, roughly 70 times the sticker price.
Here's how the tiers stack up, based on the firm's testing. The figures are the approximate API-equivalent value if you push a plan to its absolute limit (see chart above).
In other words, a maxed-out $200 ChatGPT Pro plan could represent around $14,000 of compute a month if you bought the same volume through the API — a roughly 70x markup in your favor. Claude's top tier lands near $8,000, about 40x. Even the entry-level $20 plans can return many times their price in raw usage.
One important caveat to keep in mind: these are API-equivalent figures, not what the companies actually spend. It's essentially the retail price you'd pay to buy that many tokens through the API, which is marked up well above the providers' true compute cost. So OpenAI isn't literally losing $14,000 on you, of course, but the directional point, that heavy subscribers are heavily subsidized, holds up.
Why this is happening
Every request you send to an AI model is measured in tokens, which are small chunks of text, sometimes a whole word, sometimes part of one. The more a model reads and writes, the more computing power it burns.
For instance, a quick question and a short answer might be a few hundred tokens. But the way people increasingly use these tools is far more demanding. When you hand an AI an agent-style job — plan the steps, search files, run a tool, check the output, fix mistakes, try again — it can chew through enormous amounts of text behind the scenes.
And, that hidden work is what drives costs up, and the same one SemiAnalysis researchers used to push the plans to their limits. The result is that the subscription you pay for is fixed at $20 or $200, but the cost of serving you is not. For a power user, that can balloon way past what you're actually paying.
So why do the AI companies allow it?
Because most subscribers don't come anywhere near those ceilings. The plans only tip into money-losing territory past a surprisingly low usage threshold, and the typical user stays well under it.
According to SemiAnalysis, OpenAI starts losing money on its cheaper plans once a user crosses roughly 11.4% of the allowance, and on its priciest tier, the margin vanishes at around 5.7%. Anthropic appears better cushioned, breaking even closer to 20% on lower tiers and about 10% at the top.
The blunt translation of these figures is you don't have to obsessively max out a plan to become unprofitable for the company serving you. A genuinely heavy user can get there without really trying.
For now, the labs seem willing to eat that cost. Cheap, generous subscriptions hook users, build daily habits and grab market share while the AI race is still wide open. But it's a subsidy, not a sustainable price, and subsidies have a way of ending.
The first crack is already here (almost)
You don't have to look far for evidence that the squeeze is starting. Anthropic's newest high-end model, Claude Fable 5, was reportedly bundled into its Pro, Max, Team and Enterprise plans only through June 22, 2026. After that date, access was set to move to metered usage credits unless extra capacity lets it return to the flat-rate plans. However, Anthropic suspended that model days after launch, so we don't truly know if that would have happened.
Yet, the economics behind the scenes tell an even more fragile story. Fable 5 carried an API list price of about $10 per million input tokens and $50 per million output tokens, roughly double the rates of Anthropic's already-capable Opus 4.8. Giving away that level of heavy-duty reasoning inside a standard, all-you-can-eat subscription is a massive financial drain. It highlights the ultimate irony of the "slop" crisis: while the research proves AI needs pristine, expensive data and deeper reasoning loops to avoid collapsing into nonsense, the eye-watering cost of running those premium models means the average internet user won't get to use them for free.
Even before the U.S. government abruptly pulled the model over national security and export concerns, Anthropic was already preparing to yank Fable 5 from basic subscription tiers. The high-quality web is getting pricier to build, and even costlier to sustain.
What this means for your subscription
Don't panic-cancel anything. If you're a heavy user, the rational move right now is to enjoy the discount while it's here. But it's worth bracing for a few shifts over the next year or two:
- Smarter routing behind the scenes. Expect more systems that quietly send easy requests to cheaper, smaller models and reserve the expensive frontier models for genuinely hard tasks. Done well, you may not notice.
- Usage-based pricing for the newest features. The latest, most powerful models cost the most to run. Rather than bundling them into a flat fee, companies may increasingly meter them by consumption, closer to how API billing already works, and exactly what the Fable 5 change looks like.
- Tiering of the best stuff. Cutting-edge capabilities could land behind higher-priced or pay-per-use options, while everyday chat stays cheap.
Final thoughts
If your AI subscription feels almost too good for the price, that's because, at least for heavy users, it really is. This research highlights the numbers on a subsidy the industry has mostly kept quiet about, and Anthropic's looming Fable 5 change shows the correction has already begun.
The era of unlimited frontier AI for a flat $20 or $200 isn't over yet, but the economics are pulling hard in the other direction. My advice is lock in the value while it lasts.
Follow Amanda Caswell and stay ahead of the AI curve 
Follow Tom's Guide on Google News and add us as a preferred source to get our up-to-date news, analysis, and reviews in your feeds. Subscribe to Tom's Guide on YouTube and follow us on TikTok. Finally, you can visit our dedicated Tom's Guide Savings Squad hub for expert help on getting the best products for less.
More from Tom's Guide
