I interviewed 3 YC founders on pricing AI products (they all hate usage-based)

[cornbreadfed614]

There's an "AI changes everything" narrative that makes it look like we need to reinvent every single thing about the software business.

One of those things is pricing: People on social media love to say that pricing needs to change for a few reasons:

a) Companies are cutting costs and don't want yet another subscription. Companies should just pay for what they use.

b) With LLMs/Image generators, you pay for every user interaction. Pricing should reflect that. Otherwise, a single power user can bankrupt you.

Everyone seems to be saying this. Well, everyone but people who actually price AI products lol. I recently interviewed the founders of Kraftul (YC S19), Infer (YC S19) and Ellipsis (YC W24).

A few highlights:

From Yana Welinder (Kraftful): “We've considered usage-based pricing but found its unpredictability deterred potential customers, with some exceeding their budgets unintentionally. This negative experience led us to seek a more predictable and customer-friendly model.”

=> It's easy to think that usage-based is always more customer-friendly. It's not: A lot of the time (esp. in b2b) customers would prefer to know "how much is it" than to have to constantly monitor their usage. The customer's time is also a cost to them.

From Vaibhax Saxena (Infer): “[Usage-based pricing] doesn’t help us by any means. We want to be certain of a minimum revenue every month. We do not work with customers who have 10 calls a day. People invest a lot of time to set it up, and that’s not worth it. We can just get a customer who has 30-40 calls a day.”

=> This is another important part. Your pricing needs to serve your company. If your company fails because your pricing couldn't keep it afloat, that's way worse for your customers than paying slightly more.

From Nick Bradford (Ellipsis): “We found usage-based to be more popular with hobbyists, students, and open source, because their usage was often unpredictable. In a larger company, the decision-maker has a set budget and needs to decide how much they can allocate to your product. This tends to be fine because larger organizations also have more predictable usage.”

=> I love this insight. Pricing not only depends on how you like to sell, but also on how your customers buy. If a customer wants your product, you should make it easy for them to buy. Inside companies, it's hard to get a spend approved on an expense of "whatever it costs" vs. "x$/mo".

If you're curious, I published the full interviews and my takeaways here: https://www.commandbar.com/blog/usage-based-pricing-yc-founder-interview/

 

[aprilranta]

@cornbreadfed614 Great content

 

[cornbreadfed614]

@aprilranta Thank you!

 

[hbwink]

@cornbreadfed614 Thank you! I have the belief that eventually ChatGPT will have a ‘credit card’ of sorts - where I enter my card for the website, it uses compute assigned to me via my gpt account, and I am charged through my gpt account - with openai acting as the bank

 

[agentsmith]

@cornbreadfed614 Ah shit! Very helpful post OP, thank you so much!

 

[cornbreadfed614]

@agentsmith Thank you!

 

[christie_navarro1414]

@cornbreadfed614 How much did they price their products?

 

[cornbreadfed614]

@christie_navarro1414 Kraftful's pricing is public afaik!

 

[buchanandoug]

@cornbreadfed614 Makes a lot of sense. How do you model out the monthly subscription cost without usage data though?

 

[cornbreadfed614]

@buchanandoug I mean in the beginning you're always guessing. But you can also model it slightly depending on your product.

If you have a product like Ellipsis (AI code review), you can estimate that even the most productive engineer will never commit 2000 times in a single day and bankrupt you.

It's different for products where some users might use it exponentially more than others, i.e. if you do AI content generation, you could easily have someone use it to generate a million words in a day. That's more of an issue.

You can always start with rate limits as an insurance policy and if users constantly run into them, see if you need to raise them.

Or start with a group of test users, see how their usage is distributed and model your plans on that.

 

[buchanandoug]

@cornbreadfed614 Good ideas, thank you

 

[ah44257]

@cornbreadfed614 I can't see your comments

 

[meki]

@cornbreadfed614 One of the pieces of feedback we got from retailers is they HATE %OF Revenue deals. They cannot scale when everyone is scaling with them

They much prefer $1200/month whether they run $1mm/month or $10mm/month over it

But smaller merchants want pay as you go. So we do both.

 

[cornbreadfed614]

@meki yup, best data on this stuff always comes from the people you sell to.

https://preview.redd.it/thjynzxn48y...bp&s=8d18ffe883fba9c037b5c098587d4729a969bb23

 

[meki]

@cornbreadfed614 100%.

 

[dharmmy]

@cornbreadfed614

its unpredictability deterred potential customers, with some exceeding their budgets unintentionally

This is the exact same fear the provider/developer has, they are just passing it on to the customer so they don't have to worry about it. I mean, the developer is worried about how many tokens their customers will cost them with OpenAI, but it seems kind of silly to a) recognise it as a problem (for yourself) and then b) think "screw it, our customers won't mind it".

 

[cassandra]

@dharmmy I'm going with usage based pricing for my product that is both B2C web app & B2B API. Not because I want to pass any inconvenience to my users, but because it's the most transparent pricing and lets my team just focus on building without agonizing over optimizing subscription plans.

If it comes down to losing a customer over it, even after having the conversation with them, we will offer a custom subscription tailored for their use case.

 

[dharmmy]

@cassandra It’s not really about a lack of transparency. It’s a lack of predictability.

I mean, apologies if you don’t use an LLM, my point was about people using say, the OpenAI or Anthropic APIs.

It’s transparent to say “I’ll charge you $x per token” but it’s not predictable because the customer doesn’t know how many tokens the LLM is going to spit out. They want to avoid a shock at the end of the month, basically.

 

[bobbyz]

@cornbreadfed614 Interesting read. Thank you for sharing.

 

[yadav123]

@cornbreadfed614 It depends a lot on the kind of customer.

If the majority of customers are all within say an order of magnitude of each other (primarily humans using it manually), or can be rate limited to such a range, flat pricing is better.

Customers get predictable costs, and also the vendor won't be impacted too much by outliers with much higher consumption.

But if consumption can vary more than that, and you can't rate limit a given customer, you need to charge based on consumption.

Otherwise an outlier using automation can generate a million times more load than the average human user and will bankrupt you as provider while paying the same 20 bucks like everyone else.

That's exactly what happens with cloud providers, utilities providers, and why LLM API keys have no flat fee plans.