Anthropic's IPO filing shows how fast AI bills grow. What to ask your vendors

What happened

On Oct. 26, Anthropic, the company behind the Claude AI models, made its IPO filing (Form S-1) public with the SEC. It plans to list on Nasdaq under the symbol ANT. The offering could come as early as November, as reported by CBS News.

Most coverage is about the headline numbers. Revenue was $41 million in 2023, $386 million in 2024 and $4.6 billion in 2025. In the first six months of 2026 it was $16.3 billion, up 1,300% from a year earlier. Anthropic itself says it doesn't expect to keep growing at that pace.

Anthropic's revenue

From the company's S-1 filing, Oct. 26, 2026. Billions of dollars.

1. 2023Full year$0.04B
2. 2024Full year$0.39B
3. 2025Full year$4.6B
4. 2026First six months only$16.3B

Source: Anthropic, PBC, Form S-1, filed with the SEC Oct. 26, 2026.

For a company that buys AI rather than invests in it, the more useful parts are deeper in the filing. Risk factors are where a company has to say plainly what could go wrong. Several of them describe things that land on the customer.

What it tells a buyer

AI bills grow with use. Anthropic reports net revenue retention of about 900% as of Sept. 30. In plain terms, by its own measure, customers who were paying a year ago are now paying about nine times as much. It also reports more than 11,900 customers spending $100,000 or more a year, and more than 1,900 spending $1 million or more. The filing notes these metrics haven't been checked by an outside party. Still, the direction is clear: once AI is in daily work, usage climbs, and so does the bill.

Limits and pricing can change. Anthropic writes that it has changed, and may change again, "usage limits, rate limits, or model and product capabilities." It also says it may have to offer other pricing models, such as outcome or task-based pricing. It warns that retiring older models can cause customer disputes. If a workflow depends on one model at one price, plan for that to change.

Agents cost more. The filing says its agent products are "structurally more compute-intensive" than its other offerings. An agent that reads files, opens apps and works through steps on its own uses far more than a chat answer. That's worth knowing before you hand a task to an agent and get a monthly bill.

Agents need access, and access is risk. Anthropic describes Claude Cowork as reading files, navigating applications and running multi-step work on a user's desktop. The filing says tools like Cowork and Claude Code need broad permissions and deep connections to internal tools and data, and that this raises the stakes of any security incident, unauthorized action or system error. It also expects enterprise customers to require detailed logs of each action an agent takes. That's the vendor saying it, not a critic.

The vendor is coming to mid-size companies. Anthropic says it launched an AI services company, Ode with Anthropic, with private equity firms "to bring Claude's capabilities to mid-sized companies across sectors." It also sells through Accenture, Deloitte, KPMG and PwC. Expect more pitches aimed at companies your size.

When the vendor's own filing says agents need broad access, believe it and plan for it.

Where the other side has a point

• Costs per task are falling. Anthropic says Claude Opus 5.5 cut customers' cost per task by about 40% compared with Claude Opus 5. A bill can grow because people use AI for more work, and that can be a good sign if the work is worth it.
• The scary loss is mostly accounting. The first-half net loss was $80.7 billion, but $72.4 billion of it was a change in the fair value of convertible notes and similar instruments, tied to how Anthropic raised money. The operating loss was $693 million, down from $2.1 billion a year earlier. This isn't a vendor about to vanish next month.
• The 900% needs context. HackerNoon's analysis points out that strong software companies usually report around 120%, and that usage-based revenue can shrink as fast as it grows. Anthropic's filing says the same: many contracts let customers cut usage or leave with little or no notice. That cuts both ways. You have leverage too.

Other analysts have also noted the filing's warning that the gap between AI models may narrow, as reported by Yahoo Finance. For buyers, cheaper good-enough models are good news, as long as you aren't locked in.

What changes for your office, and what doesn't

Nothing in the filing changes what AI is good at in a mid-size office: reading documents that vary, drafting, sorting and checking, with a person approving. What changes is how you should buy it. Large firms are already setting limits. Accenture now funds $29 per person per month of Claude use centrally, as reported by 404 Media. A 150-person company can do the same with less paperwork.

What to do this quarter

1. Put a budget and an alert on every usage-billed AI tool. Per user or per team, monthly. Find out what happens when you hit it: does it stop, slow down, or keep billing?
2. Get limits in writing. Ask each vendor how usage limits, rate limits, prices and model retirement dates can change during your contract, and how much notice you get.
3. Give agents narrow access. Start read-only. Grant one folder, one mailbox or one ERP role, not the whole drive. Require a log of every action, and keep a person approving anything that sends, posts or pays. Our guide on human-in-the-loop automation covers how.
4. Don't lock a process to one model. Ask whoever builds your automation how hard it is to switch models if price or limits change.
5. Tie spending to a result. Minutes per order, per invoice, per quote. If an AI bill grows ninefold, the work it does should grow too. See measuring office automation ROI.

The filing is a rare, honest look at where the AI business is headed: more usage, more agents, more access, and prices that can move. None of that is a reason to wait. It's a reason to buy carefully.