Welcome back to Good Better Best.

Each week, we break down real pricing, packaging, and product moves from SaaS and AI leaders and share the ideas worth stealing.

It was a big week for PricingSaaS. We have officially joined forces with Ulrik Lehrskov Schmidt and the team at Willingness to Pay! You can read the full story here.

Our mission is to build a single platform where SaaS and AI leaders can use monetization as a weapon for growth. That might mean learning through this newsletter or our podcast (check out our latest episode on Clay!), using our free data product PricingSaaS Pulse to understand the market, or working directly with the Willingness to Pay team on a transformation project.

We couldn’t be more excited for the next chapter. If you’re working on AI Monetization heading into 2027, and could use a sounding board — grab time with me here.

This week, we're digging into how to price Agents.

Everyone generally agrees that tokens aren’t the right model for application companies, and outcomes are only possible for a small sliver of the market. The question is what do you do if you’re in the middle?

I caught up with my friend Manu Mehra, who spent years at Google Cloud and now works on strategic pricing at Databricks. He recently wrote a post about agent economics, where he shared one of the most useful tools I’ve seen to navigate this challenge.

Let's get to it.

This week we observed 275+ changes. The highlights:

Check out more updates on PricingSaaS →

Right now, almost everyone agrees that tokens are the wrong thing to charge customers for. In the last week alone I’ve read two great posts breaking down why.

The first, from a16z: "Charging per token is becoming the default for AI companies. For most AI applications, it is a mistake."

The second, from Scott Woody (CEO Metronome): "Token billing is useful infrastructure and, for almost every company, a bad customer-facing pricing model."

I agree. But "don't price per token" only tells you what not to do. For most AI companies, the harder question is what to do instead.

The way I see it, there are two ends of the market (tokens on one side, outcomes on the other), and a very wide middle, which is the hard part.

Manu has been thinking about this exact problem, and developed his Agent P&L framework to solve it. First, some context.

If you sell model access, tokens are the right meter. Model providers and inference clouds sell compute, and their customers are buying compute. a16z makes this the first rung of their pricing ladder: if you sell model access, price tokens.

The problem starts when applications carry that logic up the stack.

Scott puts it bluntly: token billing positions your product as "a commodity markup on top of an actual commodity." Every time model prices fall, that markup gets harder to defend.

At the other end is outcome-based pricing.

Fin charging per resolved conversation is the go-to example. It works when the outcome is clean, countable, and clearly caused by the product.

But that's a short list.

a16z says to price outcomes only where attribution is clean. Scott goes further: "Outcome-based pricing only works in a very limited set of circumstances," and he argues even Fin is really an output that's marketed as an outcome.

That leaves almost everyone else: the AI apps and agents that sit on top of model costs, do real work, but can't cleanly attribute a business outcome.

The pricing mechanisms for this group are pretty well covered. Credits, outputs, hybrid seats plus usage.

But as a16z points out, "a credit is a currency, not a value metric." A credit system tells you how to bill. It doesn't tell you what your price should be, or how to defend the gap between your price and your token costs.

For companies in the middle, that gap is the challenge. They need to define their value, articulate it, and justify what they charge on top of tokens.

Manu has spent nearly a decade in pricing across Thermo Fisher, Google Cloud, and Databricks. In The Rise of Agent Economics, he starts from a simple observation: Companies can tell you they're burning "millions and billions of tokens," but not whether that consumption is "really driving the business value."

His answer is to give every agent a P&L:

From there, he uses three simple formulas:

Manu created the framework to help companies judge the agents they run. But flip it around, and it's one of the most useful tools I've seen for pricing in the middle.

In other words, the agent P&L lets you sell on outcomes without billing on outcomes. For the wide middle, that's the sweet spot.

To ensure the value side holds up, Manu recommends watching four metrics:

Number three hits especially hard for me. At PricingSaaS, we have built and tested tons of agents for all sorts of use cases. Some are amazing, some are frustrating, and accuracy and rework is the biggest variable that separates the two. If I'm wrestling with an agent for an hour to get something I could have done myself, it doesn't matter how cheap the tokens were.

Tokens work for models and infrastructure. Outcomes work for a small sliver of companies with clean attribution. Everyone in between has to define, articulate, and defend the value they add on top of token costs.

If you're in the middle, ask yourself:

I think the companies that win the middle will be the ones that do this math for their customers. And when outcomes do get standardized, they'll be in the best position to move up to outcome-based pricing.

Huge thanks to Manu for the conversation. If you haven't yet, read his full post here.

Thanks for reading! If you're working on AI monetization and want to learn more about how we help, book time here.

Until next time,

Rob