You might've noticed a pretty huge shift in the way the average user uses AI between 2022 and now. I'm not really talking about using it for more advanced workflows and relying on it more. That was bound to happen. In late 2022, we got our first real look into what AI was when OpenAI launched ChatGPT publicly, and now practically every tool is just an AI tool in disguise. That part of the story has been told over and over again. The shift I'm highlighting here, though, is this: if you use AI for anything more than the occasional search-replacement question, you're almost certainly paying for it now. In fact, you might be paying for two or three different tools, because each one has carved out a specific thing it's best at.
I know local LLMs exist that you can use on your own hardware for completely free, but I know well that the average user doesn't have the hardware, the patience, or honestly the interest to set one up. For most people, "using AI" means opening a chat tab in a browser, and the tab they open is one they're paying for. Three years ago, you could get by being subscribed to absolutely nothing. The free tiers were generous enough that most people never really hit a wall. Now the free tiers mostly exist as a funnel, and the actual product sits behind a paywall. Claude is the AI tool that everyone's been talking about lately, and its free version is, frankly, unusable. Its Max tiers are targeted at power users, so Claude Pro is what most people end up on. Unfortunately, while I once recommended Claude Pro to anyone looking to get an AI subscription without giving it a second thought, I'm finding it harder and harder to justify it nowadays.
Claude's paid subscriptions no longer cover programmatic use
The "free monthly bonus" that costs you money
First, let's talk about the most recent change Anthropic made to Claude paid plans. On May 13, Anthropic announced that starting June 15, all programmatic usage (anything you run through the Agent SDK, third-party apps, claude -p) will draw from a new "dedicated monthly programmatic credit pool" separate from your regular subscription. The company framed this as a free monthly bonus where you essentially get API credits on top of your plan: Pro gets $20, Max 5x gets $100, Max 20x gets $200, with per-seat allowances for Team and Enterprise.
Now, while the company might've framed it as a bonus, API pricing is more expensive than subscription pricing. A $100 Max plan typically gives you what would cost thousands of dollars a month at API rates, and that's why paying for a subscription in the first place is worth it. Up until this change, your programmatic usage ran at those same subscription-friendly rates. Now it's been pulled out into a separate pool that runs at full API pricing. So a "free $100 bonus" for a Max user isn't really a bonus at all. It's a small fraction of the value their subscription used to deliver for the exact same use case!
The credits also don't roll over, which means if you don't use them, they just disappear at the end of the month. And once you do burn through them, your programmatic usage either pauses until the next reset or gets billed at full API rates if you have usage credits enabled. So, all the above was to say, you'll soon pay pay-as-you-go prices for something your subscription used to cover.
Anthropic also stripped third-party tools back in April
Where the pattern actually started
Anthropic also stripped third-party tools like OpenClaw out of Claude subscriptions back in April. On April 4, the company emailed subscribers letting them know that they could no longer use their Claude subscription limits with third-party AI tools like OpenClaw. Previously, tools like this would draw from your Claude subscription usage limits. However, per this change, you'd need to enable Extra Usage to keep using them, which meant getting billed at full API rates for anything that used to be covered by the plan you were already paying for.
To soften the blow, Anthropic offered a one-time credit equal to your monthly subscription price at that time, plus up to 30% off usage bundles. If that playbook sounds familiar, it's because it's essentially the same move Anthropic would pull a month later with the Agent SDK change I just walked through above. The only difference here is that the change above gets you a recurring monthly credit instead of a one-time payout, which obviously sounds more generous on the surface.
Anthropic's deal with SpaceX should theoretically have eased this
A week of "more headroom," then a month of less
Boris Cherny, the Head of Claude Code, addressed the change above and said that Anthropic hadn't factored third-party tool usage into the Claude subscription plans. Therefore, this change was, in his framing, an engineering and capacity decision rather than a pricing one. I genuinely do get that running these models is expensive, and that the subscription plans the company built a year ago were never designed for how people use Claude today. That's a very real problem. But on May 7, at its Code with Claude developer conference, Anthropic announced a deal with SpaceX to use the entire compute capacity of SpaceX's Memphis data center.
Anthropic said that the deal was specifically intended to increase usage limits for Pro and Max subscribers, and the company doubled Claude Code's five-hour window limits, removed peak-hour limit reductions, and raised API limits for Opus the same day. Then, six days later, Anthropic announced the Agent SDK change I mentioned in the first heading.
While I'm not claiming that SpaceX compute should have instantly solved every capacity issue, the timing here is hard to ignore. Anthropic spent one week telling paying subscribers that compute was about to get easier, and the next week announcing a change that quietly makes the subscription cover less than it used to. You can't run both of those plays in the same month and expect people not to notice!
They've also quietly been testing what they can get away with removing
The test that "wasn't really happening"
Now, the two changes above are official and have already been confirmed by Anthropic. Unfortunately, the company has also been quietly running experiments on the Pro plan that, in any other month, would've been written off as harmless tests. On April 21, Anthropic's pricing page suddenly showed an X next to Claude Code under the Pro tier with no announcement, and the company's Head of Growth later called it a small A/B test on 2% of new prosumer signups after the backlash hit.
Existing subscribers weren't affected, and the change was reverted on the public pricing page the next day. So the public story is "nothing to see here," and the actual story is that the company is quietly measuring whether it can permanently strip Claude Code out of the $20 tier at some point down the line.
This is going to be the new normal
The thing is, these tests will likely continue. And what worries me more than the tests themselves is that this whole pattern isn't unique to Anthropic, which makes me think we're seeing the start of the new normal for AI pricing rather than one company having a bad month.
Deals
Score discounts on AI software, subscriptions, and tools
Just last week, Google quietly downgraded its own $20 Google AI Pro plan at I/O 2026, moving from a straightforward fixed-prompt-count system to a compute-based credit model. With Anthropic specifically, there's a clear pattern they've been following: strip away part of the existing plan, witness the user backlash, reset limits or bump caps as a goodwill gesture, and announce the next quiet change a week or two later when the noise has died down.
And while I do truly understand Anthropic's perspective here and the fact that their sudden boom has been a genuinely difficult engineering and capacity problem to solve, the way they've gone about solving it has consistently pushed the cost back onto the subscribers who got them here in the first place! But I guess, business is business at the end of the day.