All insights
The Market6 min read

Read Your Vendor's Prospectus as Your Risk Register

Vishal Sachar

Co-Founder & CEO of CLRT

Most enterprise buyers will meet Anthropic's S-1 as a markets story: the revenue line, the losses, the valuation. That is the investor's reading, and it is the wrong one for anyone negotiating an AI contract this quarter. A prospectus is a document a company writes under legal obligation to say what could go wrong, and for an AI vendor much of what could go wrong is a description of its customers. This filing does that unusually well. It tells investors where the growth comes from, how far its performance claims travel, and what deployment in a large organisation actually takes. Each of those passages, read from the other side of the table, is an entry the buyer's own risk register should already contain. Most registers do not, because the people who keep them were never handed this document.

11,900+1
$100K+ customers at 30 Sep 2026, up from 950+ a year earlier, per Anthropic's S-1
Anthropic S-1, SEC, 2026
~900%2
Anthropic's Net Revenue Retention, 30 Sep 2026; an unaudited internal metric, per its S-1
Anthropic S-1, SEC, 2026
~20x3
Spend of the top 5% of Claude Code users vs under $75 for non-users, Q3 2026, per Anthropic's S-1
Anthropic S-1, SEC, 2026

Start with the numbers the vendor is proudest of, because they give the clearest view of the buyer. As of 30 September 2026, Anthropic reported more than 11,900 customers paying it $100,000 or more over the trailing twelve months, up from more than 950 a year earlier, with more than 1,900 above $1 million and about 250 above $10 million. The filing calls this enterprise adoption, and it is. Read from procurement, it is roughly twelve times as many organisations crossing a serious spending threshold in a single year, and a ladder each of them is expected to climb. The company puts its Net Revenue Retention at approximately 900 percent, an internal metric it says has not been independently verified. For a seller, that is expansion. For buyers, it means the customers it already had are, in aggregate, paying many times what they paid before. None of this is a criticism. It is the business model stated plainly, and the buyer's budget is the part that grows.

FIG. 01Anthropic customers by trailing twelve month revenue, as disclosed in its S-1 filed on 26 October 2026; counts are the filing's own approximations.
01Where spend comes from

The filing is just as specific about where the expansion comes from. In the three months to 30 September 2026, Claude Enterprise users who did not use Claude Code spent less than $75 a month on average, while the top 5 percent of Enterprise users who did use it spent approximately twenty times that amount. The seat is not the cost. The cost is whatever the most active people decide to delegate to an agent, and that decision is taken at the desk, not in the contract. The prospectus describes the mechanism without embarrassment. Cost per task has fallen with recent releases, by approximately 40 percent from Opus 5 to Opus 5.5, and in its own words lower costs per task drive greater usage. After Opus 4.5 arrived in November 2025 at a lower cost per task, quarter on quarter revenue growth accelerated from 48 percent to 132 percent, which the filing attributes to several factors, lower prices among them. A cheaper unit tends to produce a larger bill.

FIG. 02Cost per task falls with recent releases, while the top 5% of Claude Code users spend about twenty times the average Enterprise user without it; both are Anthropic's own figures in its S-1.
02What does not transfer

The second passage concerns how far the performance claims travel. The filing reports that Claude's success rate on hard, open-ended engineering problems reached approximately 90 percent in September 2026, up about 50 points in nine months, on internal benchmarks. In the risk factors it adds the sentence every buyer should copy into their own file: results from those benchmarks may not be replicable by customers who evaluate the models with different tasks, data, criteria or computing environments, and may differ materially from what any particular customer experiences. That is an honest and correct disclosure. It is also a transfer of responsibility. The vendor has measured the model on the vendor's work. Whether it succeeds on yours is a question only you can answer, and answering it requires a test built from your own tasks, which most buyers do not have on the day they sign.

The third passage concerns what deployment takes. The prospectus describes Applied AI teams, which include forward-deployed engineers, working with enterprises to design and deploy custom solutions tailored to a customer's own systems and processes, and notes that enterprise deals demand greater sales, administrative and professional services resources. Read as a buyer, this is the vendor saying that the model alone does not reach production in a large organisation; people who understand both the model and the business have to sit inside it. The filing also warns that agentic systems may pursue stated objectives in ways technically consistent with their instructions but contrary to the user's actual intent. Taken together, those sentences locate the real work. It is not in the licence. It is in deciding which workflows the agent is pointed at, and in checking that it did what was meant rather than what was said.

FIG. 03Four passages quoted from Anthropic's S-1, mapped to the buyer document that should carry each. A CLRT reading, illustrative.
03The buyer's half

None of this calls for suspicion of the vendor. It calls for symmetry. The seller has written down, under legal obligation, the risks its business model creates, and the buyer's job is to write down the same risks from its own side. The growth engine becomes a budget line with an owner and a ceiling, because consumption-based fees made up 80 percent of Anthropic's revenue in the first half of 2026 and the meter runs at the speed of delegation. The replicability caveat becomes an acceptance test on the buyer's own work before anything scales. The forward-deployed engineers become a question about who, on the buyer's side, holds the judgment about where to point the system. A contract negotiated without those three entries is negotiated on the vendor's numbers. The prospectus is public, free and precise. It is the cheapest due diligence an AI buyer will ever be offered.

The prospectus says it plainly: lower costs per task drive greater usage. For the buyer, a cheaper unit is how a larger bill arrives.

A deeper dive

The mechanism worth understanding is consumption pricing, because it changes who sets the budget. Under a seat licence, the buyer fixes the spend at signature. Under consumption, the buyer agrees a price and the users set the volume, one delegated task at a time. The S-1 is clear that this flexibility runs both ways: many customer contracts allow customers to reduce usage or terminate with limited or no notice, and customers may cut or stop usage at any point because of budget or procurement constraints. That is genuine leverage for the buyer, and almost nobody uses it, because using it requires knowing which usage is producing value and which is simply happening. The filing also records that about 79 percent of Anthropic's third quarter revenue came from outside its top five customers, up from about 64 percent in the first quarter. Expansion is now broad, which means it is under way in thousands of organisations whose finance teams see the invoice before anyone has asked what the tokens bought.

The second-order trap is that every improvement the vendor ships makes the register harder to keep. Cost per task falls with successive releases, in the filing's examples; Fable 5.1 cut the cost of typical workloads by approximately 25 percent, with savings of up to 45 percent on complex coding and highly agentic tasks. Those are real savings per task. They also lower the threshold at which a team decides a workflow is worth delegating, so work that was never in scope arrives in the bill without a decision ever being taken. A procurement team that benchmarks the unit price will report success while the total rises. A governance team that leans on the vendor's success rates will approve workflows never tested on its own data. The disciplined response is unglamorous: an inventory of what agents are pointed at, a test that defines good enough on the buyer's own tasks, and a named owner for every line of spend. None of that is in the vendor's gift. It is the buyer's half of the contract.

Work with CLRT

CLRT does the buyer's half of this work. We read the vendor's numbers against your operation, decide which workflows deserve an agent and which will only grow the invoice, and build the tests that show whether a model succeeds on your tasks rather than on its maker's. If you are negotiating an enterprise AI contract this quarter, start with the question the prospectus leaves to you: where should this be pointed. CLRT Ascent runs that diagnostic at ascent.clrtstudio.com, and our team will take you from its answer to a system you can govern.

Vishal Sachar

Vishal Sachar is the Co-Founder and CEO of CLRT in Dubai. He helps businesses put agentic AI to work, leads practical AI workshops for leadership teams and operators, and writes on agentic systems, AI governance and the economics of automation. Reach him at vishal@clrtstudio.com or on LinkedIn.

Start here

Skip the reading. See where your leverage leaks.

Ascent is our free diagnostic. Ten minutes, and you have the one workflow worth building first.