Are You Actually Saving Money With Claude?

Issue #111 - October 23, 2026

Author

Atiba de Souza
October 23, 2026

Have you ever upgraded to a newer AI model because it promised to be faster, better, and cheaper...

Only to wonder why you're still burning through your usage so quickly?

Because something happened this week that made me question how we're being sold AI.

And I think you should know about it.

Anthropic announced that its new Claude Sonnet 5.5 is 30% cheaper.

Sounds great, right?

I mean, who doesn't want better AI for less money?

Especially when it feels like every new model comes with a bigger price tag.

But here's the thing.

We actually tested it.

And what we found was VERY different from what that 30% number might lead you to expect.

First, you know I love Claude.

If you've been following me for a while, you know I've been pretty bullish on Claude.

I use it.

We build with it.

And I still think it's one of the best models out there for thinking, planning, and solving complicated problems.

So when Anthropic announced Opus 5.5 last week, followed by Sonnet 5.5 this week, I was interested.

Better performance?

Great.

Faster responses?

Even better.

And now cheaper?

Sign me up. 😂

But I've been around technology long enough to know something.

What a company advertises and what you actually experience aren't always the same thing.

So instead of just accepting the announcement, we decided to look at the numbers ourselves.

Here's something most people don't know about how we work.

We've built a lot of our own AI tools.

And every time one of those tools sends a request to an AI model, we record what happens.

The request.

The model used.

The response.

And, importantly, what it cost us.

Why?

Because I want to know what we're actually paying for.

And that gave us an interesting opportunity.

When Sonnet 5.5 came out, we didn't have to guess whether it was cheaper.

We could take real requests we'd previously sent to Sonnet 4.6, run them through Sonnet 5.5, and compare the results.

So that's exactly what we did.

Just over 1,000 real prompts.

Same work.

Different model.

And we looked at both cost and quality.

Now remember...

Anthropic's announcement said 30% cheaper.

Want to know what we actually saw?

7%.

Yep.

Seven percent cheaper.

Not 30%.

Seven.

Now, don't get me wrong.

I'll take a 7% savings.

I'm not complaining about saving money.

But if you told me I was going to save 30%, and I actually saved 7%, I'm going to have some questions.

Wouldn't you?

And here's where it gets interesting.

The price per token isn't the whole story.

When AI companies talk about pricing, they're often talking about how much you pay per token.

Think of tokens as the pieces of text the AI processes and generates.

And yes, the advertised price per token can go down.

But what happens if the newer model starts using or returning more tokens to complete the same task?

You might be paying less for each token...

While paying for more tokens overall.

That's an important distinction.

In our testing, changes in token usage affected the actual savings.

And when we ran a similar comparison between Opus 5 and Opus 5.5?

We got an even bigger surprise.

Opus 5.5 was 11% MORE expensive for the requests we tested.

Let that sink in.

A newer model being promoted with lower pricing ended up costing us more for the work we were actually doing.

Now, that doesn't mean Anthropic's published token prices are necessarily wrong.

And it doesn't mean everyone will get the same results we did.

Our tests reflect our own workflows.

Your results could be different.

But that's exactly my point.

The advertised price tells you what a token costs. It doesn't necessarily tell you what your work will cost.

And I think a lot of people are missing that.

This matters even if you're not building AI software.

Maybe you're thinking...

Atiba, I'm not running a thousand API requests. I just use Claude to help me with work.

Fair enough.

But let me ask you something.

Have you ever been working inside Claude, getting into a good rhythm, and suddenly...

You've hit your usage limit?

Now you have to wait.

Or upgrade.

Or change what you're doing.

And you're thinking:

I thought this thing was supposed to be getting better.

Well, that's part of the cost too.

Even if you're paying a flat monthly subscription, you're still dealing with limits on how much you can use.

So when a company announces a cheaper model, that doesn't automatically mean you're going to get more work done before hitting those limits.

You have to look at your own experience.

Are you getting through more work?

Are you reaching your limits sooner?

Are the results better?

Are you spending less time correcting the output?

Because saving money on a model doesn't mean much if you're spending more time fixing what it gives you.

And a model that costs a little more might still be worth it if it produces substantially better work.

That's why we measure quality alongside cost.

Here's the lesson I'm taking from this.

I still love Claude.

And I'm not telling you to stop using it.

The newer models are genuinely impressive in a lot of ways.

But I am becoming more skeptical of AI announcements that promise dramatic savings.

Because it's easy to get excited about a number.

30% cheaper!

25% faster!

Twice as powerful!

Those numbers sound wonderful.

But the number that matters most is the one that shows up when you actually use the tool.

I've said before that we need to stop getting attached to individual AI models.

I think we also need to stop getting attached to their marketing claims.

Test the model against the work you actually do.

If you're building with APIs, compare the cost of real requests.

If you're using a subscription, pay attention to your usage limits and how much work you're getting done.

And in both cases, look at the quality of what comes back.

Because the cheapest AI isn't necessarily the one with the lowest advertised price.

It's the one that helps you get the work done well, without wasting money or time.

That's the number I care about.

And it's the number I think you should care about too.

Can I ask you something? 🤔

I'm curious whether you're experiencing this with Claude.

For me, it was seeing 30% cheaper in the announcement, then seeing only 7% savings when we tested our actual work.

That was a pretty big eye-opener.

Have you noticed Claude getting more expensive or hitting your usage limits sooner?

Hit reply and tell me in one sentence.

Even "Yes, I'm hitting my limits faster" or "No, it's been about the same" is perfect.

I genuinely want to know what you're seeing.

Want the full breakdown? 👇

I recorded a video explaining how we tested the new Claude models, why the advertised price reductions didn't match our actual costs, and what I think you should be watching before switching models.

For the physicians reading this 👨‍⚕️

Our AiM (AI & Marketing) Rounds community has been growing, and the conversations are getting better because more doctors are sharing what they’re actually trying, questioning, and figuring out in real time.

We’re talking about practical AI, marketing, what’s working, what isn’t, and how all of this actually fits into a medical practice.

No pressure to be an expert. No need to have the perfect answer.

Just doctors learning from doctors.

If you want in, reply “JOIN AIM” and I’ll send you the WhatsApp link.

A few things worth checking out 👀

Just sharing a few things I've been exploring lately:

Was this email helpful? 💬

I'd love your quick feedback.

❤️ Yes — I needed to hear this.
💛 Interesting — I'll pay closer attention to my AI usage.
💜 I haven't noticed a difference yet.

Just hit reply with ❤️,💛, or 💜.

I read every response, and it helps me know what I should write about next.

— Atiba

Reply

or to participate.