Growth Unhinged is proudly supported by Metronome, a Stripe product

I recently attended Monetize ‘26, a conference hosted by Metronome (a Stripe product) that's becoming the go-to event on modern pricing, billing, and monetization strategy. The topic on everyone's mind: AI is forcing businesses of all types to rethink pricing. But what does that mean for SaaS businesses in particular?

HubSpot is in the midst of this transformation right now as a public company. Some HubSpot agents are now priced by outcome, not by seat.

Metronome’s Scott Woody sat down with HubSpot's Matt Clamp to discuss how HubSpot got there, why they waited to charge for AI, and what outcome pricing changed for customers and billing. Watch the fireside chat here.

I don’t know about you, but I’ve finally stopped seeing any bragging about #tokenmaxxing in my social media feed. I’ll take that as a win.

The pendulum has swung from AI adoption to AI ROI, which has me (again) thinking about AI pricing models.

The latest data from Peter Walker at OpenRouter shows that price per token has dropped by 80% since January 2025. It’s down more than 50% just in the past 12 months. Open weight models, which typically cost a fraction of the frontier, are surging.

Customers are voting with their usage, shifting to lower-cost and open-weight models. Software companies are left scrambling to respond:

  • How frequently do you need to change prices?

  • Should you slash AI credit prices when costs come down?

  • What about when a new frontier model comes out at an even higher price?

Meanwhile, Anthropic seems to be betting that software companies (and customers) will prefer to decouple AI spend from their platforms.

They launched linked billing in beta on November 17th with partners like Asana, Gumloop, Higgsfield AI, Linear, and Ramp. It allows third-party platforms to let their end customers pay for Claude API usage without sharing an API key. End customers draw down from their existing Anthropic commitment and can set their own spend limits for each platform.

These dynamics make pass-through pricing, a form of cost-plus pricing that gives customers control over their AI spend, an increasingly compelling option. Today I’ll explore the rise of pass-through pricing for AI and how it works in practice with a real-life case study from an early adopter (Gumloop).

The rise of pass-through pricing for AI

In March Clay formally split their pricing into two axes, data (passed-through close to cost) and platform actions (higher margin). Data credit costs dropped by 50-90% and Clay passed through the cost of AI models with a 0% markup.

The idea was to decouple Clay’s platform value from the data needed to fully realize this value. In this way it reminded me of Costco’s business model: charge for the membership (high margin) and then make the goods you buy at Costco as affordable as possible.

What feels different now is that more companies are embracing a largely pass-through pricing model, charging a percentage of token spend managed through the platform. This is similar to classic marketplace pricing, which charges a take rate on platform purchases. Such take rates usually range from ~10% (Etsy, Substack) to upwards of 20% (DoorDash, Uber, StubHub).

At OpenRouter, probably the most prominent example of pass-through pricing for AI, take rates start at 5.5% (Standard plan) and go up to 8% (Business plan). Inference is billed at the provider’s list price on every plan.

OpenRouter uses a credit system to operationalize this. Customers buy credits (either manually or with auto top-up) and then each request deducts credits from the user’s balance.

The model is also friendly to Bring Your Own Keys (BYOK), which gives customers flexibility to spend tokens they’ve already bought elsewhere. OpenRouter lets customers spend up to $25k of list price inference per month with no fees, then charges a 5% fee after (based on what the same request would cost at list price on OpenRouter).

While I’m not sold on pass-through pricing as the answer to AI monetization, it does have its advantages, especially when customers are fixated on reducing their AI bills:

  • It creates trust through transparency. Most credit models can feel like a black box. Pass-through pricing models clearly explain the bill so customers know what they’re paying for.

  • It gives customers more control over their bill. They have all the tools they need to save money including switching to cheaper open weight models or bringing their own keys.

  • It immediately adapts as prices change. There’s no need to rethink pricing strategy when a new frontier model comes out since pass-through pricing already accounts for it.

  • It attaches monetization to a metric that’s growing fast. As token consumption continues to grow exponentially, customers automatically spend more in a pass-through pricing model even if they feel like they’re saving money.

Inside Gumloop’s new pass-through pricing model

These factors made pass-through pricing attractive to Gumloop, the Benchmark-backed Series B automation startup.

In August Gumloop announced what they called “a simple, transparent pricing model.” In the new pricing model, Gumloop passes through tokens and compute at cost. They charge an 8% orchestration fee on top of these costs. (Gumloop also charges a separate enterprise platform fee.)

Gumloop’s logic is hard to argue with. If AI models drop in price, customers automatically benefit from it (keeping them from considering a competitor). If AI models raise prices, Gumloop doesn’t face a margin crisis.

The decision was made by the founders because the previously opaque pricing always gave prospects an opening to say they could build it cheaper. The founders wanted to price Gumloop like infrastructure, the way Stripe and Snowflake are.

Gabriela Suchanek, Gumloop’s head of finance and operations, told me Gumloop made the decision in less than a month. She synthesized feedback from customers, then ran a bunch of analysis around what would happen under different pricing models. Then it took about eight weeks to fully operationalize (this was the bigger lift).

“The transition required significant coordination across teams,” Gabriela said. “It drastically changed how we worked. For deals in the pipeline that we already sent quotes to, we’d have to send a new quote and re-explain the pricing. The operational overhead lasted until October.”

Why an orchestration fee of 8%, I asked? Gumloop ultimately settled on 8% as a balance between making AI usage affordable and capturing the value of its orchestration platform.

OpenRouter’s platform fee starts at 5.5%, which felt too low relative to what Gumloop had built for customers. But a fee of 12% or higher felt too hard to justify and could hurt usage of the platform. 8% matches what OpenRouter charges for its Business tier.

Historically, Gumloop had different pricing models for self-serve and enterprise customers. Self-serve pricing prioritized accessibility and adoption, while enterprise pricing reflected larger commitments and different customer requirements. Pricing was credit-based and customers could save by committing to more credits upfront.

The credit system felt opaque to customers who didn’t know what a credit meant to them or what to do with it (sounds familiar!). “A task could mean automating something on your Gmail or sending a lead or creating an entire workflow for lead automation,” Gabriela told me. “We had so much different feedback from customers all saying they’d like a more transparent pricing model.”

In the new pricing, every chat within Gumloop includes an exact cost breakdown: how many credits were spent on infrastructure, inference, and so on. This gave people the transparency they needed to push automations into production, and it changed the way Gumloop talked about pricing with customers. “If a chat costs 300 credits and 200 of that is inference, you can swap the model and take that 200 down to 50. For us that's a core value more than a pricing tactic.”

  • Gumloop’s price per credit dropped 70% in the new model.

  • Gumloop also launched open weight models, which made the platform even more accessible to different use cases that would’ve been too expensive for companies to run. (Open weight models now make up 20% of usage, up from 0% a few months ago.)

Usage skyrocketed, according to Gabriela who shared this chart with me. Gumloop’s Pro trial conversion rate went up 42%.

Even with usage growth, it’s likely that the financial impact varies greatly by customer based on their consumption and use cases. Gumloop will keep current contracts until renewal, then switch customers onto the new pricing. “A core value at Gumloop is to reduce the price of the credit, and we’re betting that people will use it more in the future.”

I appreciated the transparency, yet I was curious about how customers could possibly predict costs in this model. To address that Gumloop built a credit consumption prediction based on the last 30, 60, and 90 days of activity. They give this to IT managers (the Gumloop admins).

For new customers, Gumloop embraces a land-and-expand sales motion. Customers start with a specific use case, run a pilot to see value, and can predict how many credits they’d use within 12 months. “But after four months they’ve expanded to five or ten use cases and most of the enterprises are re-buying credits.”

Gabriela is constantly watching Gong calls with AEs to get a sense of how customers feel about the new pricing. She reviews these on both a weekly and monthly cadence to see what’s actually happening.

Investors haven’t pushed back about this new pricing. Gumloop is bringing them along by showing usage trends before and after the change. A favorite metric: how many months it takes enterprise customers to renew their credits.

Thanks for reading Growth Unhinged! To receive new posts and support my work, please subscribe or share this post.

When to consider pass-through pricing

I’m not convinced that pass-through pricing is the best AI monetization strategy for most companies, but it is an intriguing direction for some.

Pass-through pricing risks positioning your product as a wrapper around AI models rather than having its own standalone value. It makes it difficult to profitably serve larger customers who may require significant upfront investment and professional services, yet consume relatively few tokens at the outset.

What’s worse, it anchors your value capture to other products you don’t control. You’re in a position where your prices (and revenue) rise and fall based on the whim of frontier models’ pricing decisions.

Seriously consider pass-through pricing only when these conditions are met:

  • Your value proposition (and costs) cannot be separated from AI token consumption

  • Your product is a platform where customers see value in multiple different ways

  • Your customers grow their token consumption much faster if and when costs decrease

  • Your customers demand flexibility and choice in AI models (frontier, open weight, bring your own key)

  • Your investors are OK with revenue variability and seasonality tied to token spend

Otherwise, charge for what makes you different and pass through what doesn’t.

A platform plus token pricing model lets you do both. The bet is that transparent AI usage pricing wins customers and drives adoption, and the platform fees help you tell the story of what you’ve built above and beyond the LLM.

Related resources:

Reply

Avatar

or to participate