Frameworks, core principles and top case studies for SaaS pricing, learnt and refined over 28+ years of SaaS-monetization experience.
Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.
.png)
Picture a junior lawyer who needs to review a thirty-page contract for a deal. In the past, that might have taken an evening or two. With an agent, it becomes a short series of steps the model runs through for them. Each step gives the model a different amount of text to read and asks it to write a different amount back.
First, the agent reads the whole contract (about 12,000 tokens of input) and writes a short summary of around 800 tokens. This covers the sections, key terms, and schedules. Next comes the hard part. The agent goes clause by clause and checks each one against a list of common risks, like liability caps, intellectual property, and termination rights. This step reads about 8,000 tokens and writes the most of any step: around 3,500 tokens of detailed notes.
After that, it cross-checks the risky clauses against past cases and local rules, reading about 6,000 tokens and writing 1,200. Then it writes the report the lawyer will actually read, with risk scores and suggested edits, about 4,000 tokens in and 2,500 out. Finally, it reviews its own work against the original contract to catch any mistakes: about 5,000 tokens in and 1,500 out.
Add it all up and one review uses about 44,500 tokens, roughly 35,000 read and 9,500 written. So the agent reads about four times as much as it writes. That ratio matters, because models charge very different prices for reading and writing.
Now scale it up. Imagine 5,000 lawyers using the platform. That is similar to Harvey's reach across several large firms. Its trial with one firm, Allen & Overy, alone put 3,500 lawyers on the platform and produced 40,000 queries. Say each lawyer starts three reviews per working day, over 250 working days a year. That comes to 3,750,000 reviews a year. In tokens, that is about 131 billion read and 36 billion written each year, all from this one workflow. One review is cheap. Millions of them are not. At that scale, the cost per review becomes the number that matters most.
## The Spread
So how much does one review actually cost to run? The answer depends a lot on which model you pick. Walk down today's list of options and the same review ranges from about a penny to about forty-one cents.
Start with the frontier models. Claude Opus 4.6 costs five dollars per million tokens read and twenty-five dollars per million written. It is the premium choice, and that shows up in the cost. Run the full workflow on it and each review costs about forty-one cents. Across 3.75 million reviews, that is about $1.55 million a year. OpenAI's GPT-5.2 is cheaper, at $1.75 and $14.00 per million. It lands near nineteen cents a review, or about $728,000 a year.
The mid-tier models cost less. Claude Sonnet 4.6, at $3.00 and $15.00, runs about twenty-five cents a review, or roughly $928,000 a year. GPT-4o, at $2.50 and $10.00, comes in around eighteen cents, or about $684,000 a year.
The open-weight models cost far less again. Alibaba's Qwen3.5 397B, at forty cents and $2.40 per million, costs under four cents a review, about $138,000 a year. Its smaller version, Qwen3.5 Plus, at twenty-six cents and $1.56, costs about $90,000. DeepSeek's V3.2, at twenty-eight and forty-two cents, lands near $52,000. And Meta's Llama 4 Maverick, at fifteen and sixty cents per million, costs barely a penny a review, about $41,000 a year for the whole workflow.
Put the two ends side by side. The most expensive option costs about $1.55 million a year. The cheapest costs about $41,000. That is a thirty-seven-fold difference. Same task. Same words. The only thing that changed is the model.
## Why Writing Costs More Than Reading
Remember that the agent reads about four times as much as it writes. Here is the surprising part. Even though writing is only about a fifth of the work (around 9,500 of those 44,500 tokens) it drives most of the cost. On Opus, writing alone makes up about 58 percent of the cost per review. On GPT-5.2, it is closer to 68 percent.
The reason is technical, but simple. A model can read a long document more or less all at once. But it has to write its answer one word at a time, with each word building on the last. That slower, step-by-step writing costs more to produce. So providers charge three to six times more for writing than for reading.
This points to a useful tip. If you want to lower costs, look first at how much the agent writes, not how much it reads. You can ask it to be more concise, use clean structured formats, or write a short draft first and expand only the parts that need it. Each of these trims the most expensive part of the job.
## Using More Than One Model
Once you see the range of prices, a practical approach stands out. You do not have to use one model for everything. Most steps in our review (reading, pulling out details, and double-checking) are routine, and a budget model handles them well. A smaller share, the real risk analysis, is better suited to a stronger model. And a small slice of truly hard, high-stakes questions is worth sending to the best model available.
So you can route the work by difficulty. Send about 70 percent to a budget model, 20 percent to a mid-tier model, and the hardest 10 percent to a frontier model. Use Maverick, Sonnet, and Opus for those tiers, and the blended cost comes to about $369,000 a year. That is 76 percent less than running Opus for everything, and 60 percent less than running Sonnet for everything. And the strongest model is still there for the questions that need it. This kind of routing is one of the most effective ways to manage agentic AI costs. It often gets set up later, usually after someone in finance asks why the cloud bill looks the way it does.
## Where Engineering Meets the CFO
This brings us back to where we started. The difference between Opus and Maverick on this one workflow is about $1.5 million a year. And a company like Harvey does not run just one workflow. It runs many: contract review, research, drafting, and compliance. Depending on how model choices are made, the total model bill could land anywhere from $500,000 to $15 million a year. That is not a small technical detail. It is a real conversation about margins, pricing, and how the product is built.
On one side is an engineer who picked a model for good engineering reasons. It was fast, accurate, and easy to build with. In doing so, they also set the cost structure for the whole product. On the other side is a finance leader building cost forecasts on top of that choice, often without having seen it. Both are doing their jobs well. But each one's work depends on understanding the other.
This is the new overlap, and it is here to stay. With agentic AI, the line between an engineering decision and a financial one has mostly disappeared. The most useful approach is to treat model choice as a shared, ongoing decision. Put engineering and finance in the same room, look at the cost of the work together, and revisit the choice as the market changes. Reviewed regularly, model selection becomes a normal part of running the product rather than a surprise later on.
## A Note on Where Models Come From
There is one more point worth raising, gently, because it does not have a clean answer. Some of the cheapest and most capable open models, including Qwen and DeepSeek, come from labs in China. Their cost advantage looks real and likely to last, helped by lower labor costs, different ways of buying compute, and strong government support. For a company serving customers around the world, this raises questions that a simple cost comparison cannot answer.
They involve compliance, where data is allowed to flow, and what customers are comfortable with. The cost math might point one way. A cautious enterprise customer might point another. These are not purely engineering questions, and they are not purely finance questions. They are exactly the kind of decision that lives in the overlap this chapter keeps coming back to.
## The Model & Harness Is the Margin
So we end where we began, with a small choice that turns out to be a large one. For any company building agentic AI products, the choice of model is not just a technical detail. It is the first number we pressure-test at Monetizely. More than any other single decision, it sets what the product costs to run. At real scale, the gap between the most and least expensive options is measured in millions of dollars a year. And the rise of open models that perform well at a tenth or even a fortieth of the price is not a minor footnote. It is a real change in how the economics work.
The most effective teams treat this as an ongoing decision. They share it across engineering and finance, measure it, and keep adjusting it as the product grows. Handled this way, model choice and harness engineering becomes a steady source of healthy margins and a normal part of building the product.
Join companies like Zoom, DocuSign, and Twilio using our systematic pricing approach to increase revenue by 12-40% year-over-year.