On 26 October Anthropic filed a registration statement to list on the Nasdaq Global Select Market under the ticker ANT, with Goldman Sachs and Morgan Stanley named on the prospectus. It is the first frontier AI lab to file for a public listing, and the S-1 is the first time anyone outside the company and its investors has seen the economics of the model behind Claude, Claude Code, and Cowork laid out in full. For firms using these tools, as well as those running Copilot or ChatGPT, the economics are much the same. Their account managers will find this document particularly important.
The headline numbers are large even by this industry's standards. Revenue for the first half of 2026 was $16.3bn, up from $1.2bn in the same period a year earlier, and the company recorded an operating profit in the second quarter.
Net revenue retention was approximately 900% as of 30 September, revenue for the quarter to 30 June was up 1,368% year on year, and trailing twelve-month revenue per employee was more than $6.6m.
As of 30 September, Anthropic had more than 11,900 customers spending at least $100,000 a year, more than 1,900 spending upwards of $1m, and about 250 at $10m or more.
Two weeks ago we covered Accenture capping Claude at $29 a month per person, and the week before that Microsoft cutting its own Claude bill by more than a third. Those were the customer side of the AI bill. The S-1 is the vendor side, and it explains why the pressure is about to move from the buyer's spreadsheet to the buyer's contract.
What the prospectus discloses
Start with where the money comes from. TechCrunch's Tim Fernholz explains the 900% figure plainly: spending from the same set of customers in Q3 2025 grew nine times by Q3 2026. New logos have little to do with it; existing customers are the growth plan. The filing says Claude Opus 5.5 cut customer cost per task by approximately 40% compared with Opus 5, and that lower cost per task drives greater usage, which is the vendor's answer to a cap: make each unit cheaper so the total keeps climbing.
Then where does it go? TechCrunch's Marina Temkin, reading the same document, reports commitments of at least $110bn to AWS over ten years and at least $114bn to Google Cloud through 2033, and a $72.4bn non-cash charge tied to Amazon's convertible notes that accounts for most of the reported net loss. The filing does not disclose the total compute spend, leaving the figure unknown. However, it indicates that most of the expenditure after 2030 will fall between 2031 and 2045. The founders keep voting control through a separate share class, so the public will own the economics without steering them.
The product mix tells a similar story. According to TechCrunch’s Rebecca Bellan, more than half of Anthropic’s customers spending at least $100,000 annually now use both the Claude platform and one of its own applications, such as Claude Code or Cowork.
At the end of last year, that figure stood at around 30%. The third-quarter figures offer another indication of the value of these customers: the top 5% of Claude Enterprise customers using Claude Code spent around 20 times more each month than the average enterprise customer who did not use it. Anthropic’s prospectus makes its commercial strategy clear: customers who use more of its products tend to spend more and are less likely to switch providers.
The sentence that will shape your renewal
The risk factors contain the line that matters most for a buyer.
Our revenue is primarily consumption-based, and much of it is not subject to long-term contractual commitments, while our compute and infrastructure commitments generally extend over multiple years.
It describes those compute deals as involving "minimum purchase commitments, take-or-pay provisions, and other terms that obligate us to pay for resources regardless of actual utilization."
Put the two halves together. Anthropic has promised to pay for compute whether or not customers use it, and most of its customers can walk away at the end of the month. A company in that position, now answerable to public shareholders, has one obvious move, which is to get customers onto commitments too.
That looks like annual or multi-year minimum-spend agreements with discounts for committed use, usage caps for anyone who stays on pay-as-you-go, and steering towards cheaper model tiers when the frontier model is capacity constrained. The filing already notes that "we have also changed, and may in the future change, usage limits, rate limits, or model and product capabilities."
Anthropic has not announced any pricing change, and nothing in the filing says it will. We are reading the structure, and the structure points towards commitments. Microsoft's internal cut and Accenture's $29 cap were the customer response to that pressure, and they arrived before the vendor's economics were public.
Why is leaving getting harder each quarter?
Jellyfish, which makes engineering analytics software, published a report on 27 October based on agentic coding tool use by 276,000 of its customers. Claude Code and other tools built on Anthropic models accounted for nearly a third of all proposed code changes submitted for review, and spending on those tools has grown tenfold since the start of the year, to about 2% of total engineering cost.
Even if you're saying 2% is still small, that 10x growth is growth that most CFOs didn't plan for.
Nikolas Albarran, Jellyfish's AI product lead, told TechCrunch why a firm that notices that line rarely switches vendor: "Even if they're all mostly the same, they're all going to be slightly different, have their own slight quirks of how they handle agent.md files, etc., and so that cost to switch is not zero." Swap agent.md files for the skills, templates and prompt libraries your cost managers have built around one vendor's tools over the past year and the point holds.
The vendor is now selling you the meter
On 28 October, two days after the filing, Anthropic made smart reports, in-app surveys and pooled group budgets generally available for Claude Enterprise. Pooled budgets give a team one shared monthly allowance, with individual limits kept as a guardrail; when someone hits a limit, they can request more from inside Claude, and an admin gets alerts at 50%, 75%, 95% and 100% of the budget. Smart reports analyse a sample of a team's transcripts, group the work into workstreams and show the spend for each alongside the deliverables produced. Anthropic says they are not for evaluating individual performance.
Read that against Accenture's memo. The problem there was a $29 cap measuring tokens rather than work, and Anthropic has now shipped a tool that reports work. It has also shipped the tool that lets a finance director see what the spend was bought right before the renewal conversation, and that is the half to notice.
That conversation will be harder than vendors expect. Torsten Slok, chief economist at Apollo Global Management, looked in a note reported by TechCrunch on 25 October at what he calls the S&P 493, the index minus its seven largest technology companies, and found profit margins and revenue per worker flat.
AI may be a great calculator, but in business, intelligence is mostly judgement.
If the vendor's evidence is a smart report and the economist's evidence is a flat margin, the firm in the middle needs its own.
Takeaway
The first frontier lab to go public has told the market that its revenue is month-to-month and its costs run to 2045 and that its growth comes from existing customers spending more across more products. Whichever AI vendor your firm uses, that is the model they are all now measured against, and the likely response is commitment pricing. The time to decide what you are willing to commit to is before the account manager proposes it.
Pull your current AI spend across every tool and every project code into one figure, because the first thing a committed-spend offer does is anchor on last year's number.
Decide which deliverables need a frontier model and which run fine on the cheap tier, so that model steering is a choice you made rather than one made for you.
Build the private evaluation we described in the Microsoft piece, a handful of your own cost plan and programme tasks, and rerun it at every model change.
If you are offered a minimum-spend contract, ask what happens to the rate when a cheaper model arrives mid-term, because the filing says cost per task fell 40% in one model generation.
Project Flux will follow the listing through to the first trading day and what the first committed-spend AI contracts look like when they reach consultancies and contractors. Subscribe at projectflux.ai to read that before your account manager calls.
Links and Stuff
All content reflects our personal views and is not intended as professional advice or to represent any organisation.
