This is a mental discussion I have been having for the last two months. It is about pricing in AI.

I have one actionable recommendation: Kill that damn model picker.

I have been coding a specific AI app. Exciting stuff. Product is obvious. I even got a plan for distribution from day 1. Or day 0.

Everything clicks. Except one thing.

Pricing.

It's tormenting me.

Everyone describes AI apps the same way: "Cursor for X."

Fair enough. Cursor nailed something important.

They saw developers copying code from ChatGPT, pasting into VSCode, running it, fixing errors, repeating. Painful. So they brought everything together. Smart move.

All good so far.

Except pricing. Or so has it been told.

Pro comes with at least $20 of included usage per month.

So the market criticism has been:

• Cursor charges you $20 for unlimited usage (or close to it)

• Cursor wants and acquires great users and developers who would like to use more and more AI at work.

• Users love the product and start using. The more they like it, the more they use it.

• All of a sudden, each power user, Cursor’s sweethearts, become the enemy of Cursor’s board because they are the least profitable users. VCs pay the bills.

• The best friends, on the other hand, are those who downloaded Cursor, paid for it, and never use it.

Cursor’s incentives start opposing the good of their best customers.

Now here is the twist.

There is nothing wrong here.

This is nothing new.

Spotify had this problem since the start. Paying 70% of their revenue in Cost of Goods Sold to music label companies in form of royalties.

This means that for a $12 Spotify subscription, record labels are making around $8.2 on average.

For users, who have Spotify open 15-16 hours? They are costing Spotify even more in royalties.

This is not to say that Spotify is a great business that we all should follow but it is a tangible instance of a publicly traded $150 Bn market cap company. So you better damn pay attention.

Spotify has to carefully manage margins and find ways to increase it. There are several ways to do this but essentially:

1. Increase prices

   1. Acquire pricing power with branding and network effects

2. Drop costs

   1. Negotiate with labels and artists

   2. Artists need Spotify more than Spotify needs them

(This is what being a McKinsey consultant feels like; telling companies to drop costs and increase prices.)

Now for the user experience:

I use Spotify because I would like to listen to music.

Specifically I listen to Led Zeppelin, Radiohead, AC/DC, Black Sabbath, Chumbawamba and the occasional Lady Gaga (ik, shut up).

I know these artists well but I have no freaking clue who the labels they work with are.

Nor do I care. As far as I can see, Spotify does not even show the label an artist is signed up with.

The label is invisible.

Which brings me to AI apps.

Open any AI app today. What do you see?

A dropdown menu. Claude 3.5. GPT-4. Gemini Pro. Pick your model.

We build this because we're developers. We follow benchmarks. We debate on Twitter which model handles recursion better. We care about this stuff.

Last week, I had dinner with an ex-banker friend. Smart guy. I'm rambling about Claude Opus being better than o3. He stops me. "How's ChatGPT doing?" he asks. He thought o3 was a company. Claude? Maybe my ex-girlfriend.

And he goes: “How is ChatGPT?”

He did not know the models, the providers or anything else. He did not even know the picker inside ChatGPT. He thought o3 was another company, let alone him thinking Claude is my ex girlfriend.

He doesn't know models. He doesn't want to know models.

He wants his work done.

The point is: once you remove the model picker, you turn the pricing problem into a product problem. And product problems? Those you can solve. You can route queries to cheaper models when possible. You can use open-source for basic tasks. You can cache responses. You can build your own models for specific use cases.

The user doesn't know. The user doesn't care.

In addition, you make the experience simpler for the users. He does not have to select the whatever model that just got released last night and is the best one based on vibes among the 7,000 developers living in Twitter.

They just gotta get the job done.

When you do this, you become a better Spotify. You can focus on making the product purely better by:

1. Adopting small and cheap focused models (usually open source)

2. Router layers

3. Non-AI functionality

4. Developing your own models

Does the user need to know? No.

Now I know, this is not a revolutionary idea where I just solved all your problems around pricing, aligning incentives, cost and revenue drivers and more.

But here's the critical part: You have to nail the routing. Users don't know model names, but they absolutely know when outputs get worse. One bad experience from aggressive cost-cutting and they're gone.

There are many things you can do here to charge more:

1. Expose the model picker in settings and make it even more premium

2. Expose the system prompt

3. etc

Take insurance and all-you-can-eat buffets:

Insurance:

• How do you price coverage when some customers cost $0 vs some customers cost $100,000 in claims?

• They know 80% of customers will cost 0, 15% moderate, 5% heavy. The 80% subsidize the 5%.

• Then they make the product better little by little: Safe driving discounts, wellness programs, security system discounts. They actively reward behavior that reduces their costs.

Mind that insurance industry is the 6th largest sector in the world and claim to have some of the largest companies ever.

In short:

1. Remove that damn picker

2. Make the product perform better

3. Try and get to pricing power (distribution)