The AI Pricing Spectrum

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.

Two AI launches over the past week fall on opposite sides of the economic spectrum. DeepSeek's V4.1-Flash is pushing monotonous work toward free, while OpenAI's GPT-6 Astra makes Computer Use real enough to handle work a human would otherwise do. Below, I mapped out the AI pricing spectrum, and what it means for packaging AI for your customers.

Before we get there, we have two Office Hours sessions coming up:

Let’s get to it.

This week we observed 200+ changes. The highlights:

Check out more updates on PricingSaaS →

Here’s what was top of mind in Pulse this week:

🔥 Hot Companies

🚨 Hot Topics

Two models launched this week on different ends of the agent economic spectrum:

Put them together and you start to see where agent pricing is headed: the cheap end is approaching free, while the capable end is approaching labor.

These two extremes bookend the playbook taking shape around AI pricing:

We're moving from agents that generate things to agents that do things. As that happens, the economic range of agentic work is widening, from fractions of a cent for monotonous tasks to meaningful dollars for hours of human-like work.

When an agent works on a task, it has to keep track of what has happened so far. The longer it runs, the more context it carries and the more expensive it becomes.

Without getting into the technical details, DeepSeek found a way to make long-running AI work far more memory-efficient, and their newest model, V4.1-Flash, dramatically reduces that burden.

That doesn't mean V4.1-Flash can do everything frontier models can, and that’s the point. There is an enormous amount of agentic work that doesn't require the best model in the world: populating spreadsheets, updating CRM fields, processing hundreds of documents, categorizing support tickets, or enriching accounts.

For this kind of monotonous work, costs are falling toward the point where vendors can simply absorb them.

At the other end of the spectrum is GPT-6 Astra, and the concept of Computer Use.

Computer Use lets an agent operate software the way a person would: open a browser, log into applications, conduct research, build a spreadsheet, move information between systems, and complete multi-step workflows. OpenAI and Anthropic have offered versions of this for a while, but in my experience they haven't been very good.

Astra looks like an important step forward. Tae Kim calls Computer Use the fourth exponential demand wave in AI:

(1) chatbots → (2) reasoning → (3) agentic coding → (4) agentic computer use

The important thing about this end of the spectrum is that the agent isn't just generating something anymore — it's doing actual work, and that work can get expensive.

Inside OpenAI, the median researcher barely used coding agents at the beginning of the year. By mid-August, the median researcher was consuming more than $600 per day of tokens at API prices, while the 90th percentile was above $7,000.

The natural response to this widening cost spectrum is routing, which isn’t new. Most sophisticated AI products already use different models for different jobs.

What's changing is how important routing becomes to the economics of the product.

As the gap between cheap and expensive models grows, vendors have more room to optimize what happens underneath a single customer experience. A task that doesn't require much reasoning might cost fractions of a cent. A more complicated task might require a frontier model and cost 100x more. The customer doesn't necessarily need to know which model did what. They just need the work to get done.

That means model selection increasingly becomes the vendor's problem, not the customer's. The better you are at breaking work into tasks and routing each one to the cheapest model that can reliably complete it, and metering effectively, the more margin you can create without changing the product or the price.

Two companies could sell effectively the same agent for the same price, but the one with better routing and metering could have dramatically better margins.

While routing is an effective way to delegate tasks to the right model, the cost difference between cheap and frontier models is widening so much, it raises the question of whether or not there should be controls on the most expensive models to protect customers from burning through all their credits in a handful of tasks.

Credit models make a lot of sense when the underlying work has relatively similar economics. They get harder when one task costs pennies and another costs dollars.

Imagine giving customers one AI credit balance covering everything from basic enrichment to an agent operating a browser for three hours. Customers will have a hard time grasping why 100 simple tasks barely touched their balance while five advanced tasks wiped it out.

We're already seeing companies grapple with this. Notion recently told admins that its premium models, Claude Fable 5 and 5.1, would begin drawing down Notion credits.

Notion explicitly says these models are designed for more complex, longer-running work and cost substantially more to operate. So they turned access off by default. Admins decide who can use them and how much each person can spend. I think we'll see a lot more of this as the cost range of agentic work expands.

The cheap end gets bundled. The middle gets metered. The expensive end gets governed. And eventually, the most capable end may be packaged separately.

Once an agent can operate a computer for hours and complete work that previously required a person, I'm not sure credits are the right abstraction anymore. The better abstraction might be labor.

Instead of selling 50,000 AI credits, layer an Agent on top of my plan to do the work for me. Here’s a mock-up I created for Clay:

A $2,000/month digital worker with a specific job description, and clear output is easier to forecast than an AI credit balance that might disappear depending on how complicated the work becomes.

There’s also an argument that a vendor-side agent is more token efficient than the end customer, since the vendor likely knows the best way to use their product.

This also avoids forcing outcome-based pricing before customers actually want it. Agents may increasingly deliver outcomes, but that doesn't mean customers necessarily want to pay per outcome.

Buyers still care enormously about predictability. A fixed-price digital worker gives them both: work performed on their behalf and a bill they can understand.

I think this leaves SaaS companies with a playbook that’s starting to get clearer:

Trying to squeeze the full range of agent capability onto one meter probably won't make sense.

The future of agent pricing isn't one model. It's a stack: bundled automation at the bottom, metered intelligence in the middle, and digital labor at the top.

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