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.
Over the past couple weeks, listening to earnings calls and watching model releases, a few things are becoming clear about AI monetization:
SaaS and AI companies are increasingly selling the harness
Models are starting to talk about price-per-task, and
Buyers still prefer predictable pricing
These factors are changing how companies position their product, and the go-to pricing model is taking shape. Let’s get to it.
Before we get there, two Office Hours notes:
First, we had to reschedule this week’s session on Selling and Renewing 7-Figure Deals. We’re still finalizing a new day/time, but if you’d like to sign up and stay informed, you can register here.
We’ve got a great session coming up on Thursday, 9/17 where I’ll be hosting David Reid, Senior Managing Director at Teneo. David and his team have been helping some of the largest legacy SaaS products transition to AI, and we’ll be breaking down 10 Trends in AI Monetization. Register here.
🔌 PricingSaaS Partners power the next era of SaaS pricing
This Week in Pricing, Packaging, and Product
This week we observed 200+ changes. The highlights:
Stripe restructured Metronome-based usage pricing and launched a new T600 card reader at $329 [Link]
OpenAI added a $100/mo Premium seat tier to Business plans, alongside a new 200-employee cap [Link]
Vercel renamed Connect to Token Requests, cut the included quota, and raised overage pricing [Link]
Intercom ended its Essential plan's 35% new-customer promo [Link]
LastPass raised Business Max from $9 to $11 per user per month [Link]
JetBrains added multi-year billing with a 2-3 year price-freeze guarantee [Link]
Supabase dropped BYO Cloud support from its Enterprise plan [Link]
Postman added Cloud performance testing at $0.04 per virtual user hour [Link]
Chargebee renamed Enterprise Plus and moved three features to paid add-ons [Link]
Cohere added a document-parsing model with new tiers [Link]
Microsoft Teams extended its Copilot add-on discount through 12/31 [Link]
Coda swapped success planning and value reviews for priority support and live chat [Link]
Checkr added a social media search add-on at $29.99 per check [Link]
Apify slashed the Free plan's concurrent Actor runs from 25 to 5 [Link]
Workato discontinued its EDI (Orderful) and VPW add-ons [Link]
Wispr Flow ended Enterprise's hidden pricing and unveiled an $18/user/mo Growth tier [Link]
Koala AI cut its Essentials tier and repackaged Boost with a Brand DNA agent plus credits [Link]
Check out more updates on PricingSaaS →
Selling the Harness
This week, both Box and GitLab reported earnings. Different companies, different categories, but both CEOs ended up explaining the same thing: what they’re actually selling as models continue to commoditize. They landed on the same word: Harness.
First, it’s worth clarifying what harness means. Both GitLab CEO Bill Staples and Box CEO Aaron Levie both gave it their own spin in their respective calls.
Here’s Staples:
"We have several structural advantages versus typical AI native tools whose monetization is based on tokens. One of the beautiful things about Duo Agent Platform is that we provide it in a cloud-agnostic, model-agnostic way, meaning we support all of the models, including open source, open weight models. We allow customers to deploy it and use it in any cloud, including, for example, air-gapped data center environments that they run. What that means is, for many of our customers, the token or inference cost is actually not embedded in the GitLab agreement. They pay us for the access to the platform, and they pay for the work done in the platform, the context, the harness, the governance and auditability that we provide, not the inference. Those are all very high-margin products."
And Levie:
"The momentum we're seeing generally correlates to two dimensions: longer running agents that do more processing work in a single session, or the ability to run agents off of large amounts of data. Both have the tendency to be very token intensive. This is great for us, because it means that customers will increasingly want those types of workflows to happen inside of platforms that are model neutral, because the more tokens your use case requires, the more obviously over time you're going to be price sensitive because you want to make sure that you're optimizing that cost structure for the use case. By having a model neutral layer, which is our agentic harness, we can then make sure that we are directing the workload to whatever is the effectively cheapest model at the accuracy level that the customer is looking for."
Both alluded to their value not being in offering tokens, but in helping customers deploy tokens productively. In this context, the agentic harness is the orchestration layer that manages tokens and inference.
Another way to say “manages tokens and inference” is helping customers complete tasks and do jobs.
The shift to price-per-task
At the model layer, the pricing conversation is shifting towards jobs as well.
This week, Anthropic released Fable 5.1. The sticker price was the same as Fable 5, but cache prices dropped 75%, from $1.00 to $0.25 per million. This is important, given cache is used for repeatable agentic workflows.
While Anthropic didn’t drop the list price per token, making cache cheaper allows them to reduce the price-per-task. Ben Thompson from Stratechery summed it up well:
Cache is particularly important for agentic workflows, which involve a lot of tool use (i.e. actual definitions and schemas which have to be loaded into the context window and which is the same across not just a single user but all users). Cache is, particularly relative to actual generation, much closer to free (the entire Internet is, in many respects, just a massive caching layer; to put it another way, cache is what undergirds zero marginal costs). Therefore, it makes total sense that this is an easy vector for Anthropic to credibly promise lower prices: develop a better cache layer, charge less for tokens from that layer (which is dramatically cheaper), and the price-per-task is lower even if the price-per-generated-token is the same.
The through-line between the earnings calls and the Fable 5.1 launch is that post token-maxxing, the conversation has shifted to jobs and workloads at both the model and app level. It seems like that would naturally shift pricing to outcome-based models, but that has not been the case so far.
Buyers still want predictability
Okta CEO Todd McKinnon gave a great answer as to why they still charge for seats in Okta’s earnings call last week. Here’s his answer (shortened for brevity):
"The first thing everyone says is, 'Well, that's crazy. Seats are going away, and you've got to charge per agent.' That all may be true, but the way customers are using agents now and the way they want to buy is per user. One of our advantages is we're super close to the customers, and as I'm sure this is going to evolve, we'll iterate quickly and give them what they want. But for now, they like the predictability. It's helping these deals move faster. I think that's the winning formula for now."
He did say Okta is building the "scaffolding" for consumption pricing later, because agent usage is going to get heavy, but their customers are still happy to pay for seats.
Similarly, Figma recently reported earnings and illustrated how they’re monetizing AI with a seat based model. Here’s Figma CEO, Praveer Melwani:
“As a reminder, we embed AI credits in each of our seats. A key signal we wanted to understand was how credit utilization would trend as customers began paying for incremental usage. A full quarter in, we are encouraged by what we've observed. First, credits included in every seat make those seats more valuable, supporting upgrades, new team conversion, and retention. As of the end of Q2, approximately two thirds of paid customers with more than $10,000 in ARR added full seats compared to their prior renewal, which is consistent with prior quarters. Second, as customers exceed the credits built into their seats, they can purchase additional credit add-on subscriptions or enable pay as you go.”
Both examples show customers are still ready and willing to pay for seats, which challenges the narrative that customers would prefer to pay purely for outcomes. In the short term, it appears agent pricing will look a lot like legacy SaaS hybrid models, with a platform license for the harness, and a variable rate for additional tokens.
We’ve seen this shape before.
A year ago I broke down Clay's restructure, and the mechanics are the same. Clay split a single credit model into two meters:
Data Credits: Used to purchase 3rd party data or AI in Clay’s marketplace.
Actions: Used to measure orchestration, work execution, and workflow steps.
In Clay’s case, they were already investing in orchestration tools, but adding any additional cost to data credits made their rates look way more expensive than competitors. By separating the two, they decoupled platform operations from data purchasing, making it clear that they’re taking a small margin on data credits, and capturing higher margins for orchestration.
In other words, they’re charging a premium for the harness. Though Clay didn’t outwardly use that phrase when they launched (at least I didn’t see it), it’s the same idea. I expect most SaaS companies will follow suit if they aren’t already.
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