- Strategic Structuring
- Risk Mitigation
Capital Markets
- Digital Engagement
- Market Visibility
Public Perception
- Shareholders Services
- EDGAR Filings
Two features help explain this business. Customers can spend more as they give Claude more work, even without adding employees. Meanwhile, Anthropic must arrange much of the computing capacity for that work long before it knows exactly how much customers will use. That combination creates room for rapid profit improvement, but also makes the timing of growth unusually important.
About 80 percent of Anthropic’s revenue in the first half of 2026 came from usage-based charges. Customers pay according to how much service they consume. This gives the business a way to expand inside an organization beyond selling another subscription to another employee.
Rakuten provides a concrete example of the work customers can delegate. In a June 2025 case study, the Japanese technology group described engineers using Claude Code for testing, debugging and documentation, with multiple coding sessions running in parallel. It reported that engineering teams’ average time to market for new features fell from 24 working days to five, and outlined plans to widen access after successful pilots.
The case study describes Rakuten’s own experience and does not disclose its spending. It illustrates why a customer might broaden adoption: a tool that helps teams deliver useful work sooner can earn a larger role in their workflow.
For Anthropic’s usage-based business, an occasional coding question and a long assignment involving repeated tests can consume very different amounts of service. More delegation can generate more paid usage without a larger customer headcount, provided the results justify both the bill and the effort of checking them.
Some enterprise agreements include annual or multiyear minimum spending commitments. Those commitments provide a floor for the business they cover, while consumption above the floor still depends on what customers choose to do. This gives Anthropic a mixture of committed spending and room for further expansion. It also makes the quality of adoption important: a useful deployment offers a reason to keep spending after the initial agreement.
Existing accounts provide evidence of substantial expansion. For the customer group included in Anthropic’s retention measure, revenue in the three months through September 2026 was approximately nine times revenue from the same group a year earlier. The measure excludes Pro and Max subscription customers and starts with customers that generated at least $1 of revenue in the earlier period.
Figure 1. Same customer group; Q3 2025 revenue = 1. Includes customers with at least $1 of revenue in Q3 2025; excludes Pro and Max subscribers. Source: S-1, pp. 3–4 and 136.
Following the same group separates spending growth by existing customers from revenue brought in by new ones. A company can grow quickly by winning many first-time buyers while older buyers barely change their spending. Here, the included existing customers collectively generated much more revenue. The result supports the idea that an initial deployment can grow into a larger commercial relationship.
The ninefold figure describes the group as a whole. It does not mean every account increased its spending ninefold or that every customer stayed. The earlier base was also much smaller. As accounts mature, continued growth in the dollars they spend would remain useful evidence even if the percentage expansion falls substantially.
Anthropic also reports a shift in its revenue mix. Customers outside its top five accounted for about 79 percent of revenue in the third quarter of 2026, compared with 64 percent in the first quarter. The filing does not specify whether the top five were the same customers in both periods. This comparison concerns revenue concentration, while the retention measure above tracks revenue from a defined group of existing customers.
Before expanding a trial into regular work, an enterprise needs to establish which results it can trust, how much review remains necessary and whether the time saved justifies the bill. A strong demonstration and a large continuing budget are different commercial achievements. The filing’s revenue evidence shows spending has expanded; it does not measure the return earned by every customer.
A better model can lower a customer’s bill by completing the same assignment with fewer billable resources. Better technology therefore does not automatically mean more revenue for the company supplying it.
The illustration below shows two possible outcomes after the effective charge per task falls. Spending shrinks if the workload stays unchanged, but can grow if more tasks become worth doing.
Figure 2. Charge is the effective service charge per task. Hypothetical illustration, not Anthropic pricing, observed customer results or a forecast. More usage is conditional.
Anthropic’s actual charges depend on the product and resources consumed. The example isolates the business question: does the improvement create enough additional paid work to offset any reduction in the bill for existing work? A more capable model may make difficult assignments feasible. A more efficient one may make routine assignments affordable. Either can expand demand, but the additional work must actually materialize.
Anthropic describes this pattern in its own product history. Its filing says Opus 4.5, launched in November 2025, combined greater capability with lower prices and a lower customer cost per task than Opus 4.1. Revenue growth subsequently accelerated. Anthropic attributes that growth to several factors, including new products and broader adoption, as well as improvements that enabled new uses. The sequence is consistent with cheaper, more capable models expanding demand; it does not isolate the effect of a price cut.
Revenue is only one side of this calculation. Anthropic also needs the cost of serving the expanded workload to leave room for other expenses. A lower customer bill could accompany a larger reduction in delivery cost, improving profit per task. Alternatively, more revenue could require so much additional computing expense that little profit remains. Revenue growth alone cannot distinguish those outcomes.
Existing accounts have already become larger sources of revenue. Sustaining that expansion will require products that keep earning a place in customers’ work, and computing resources that let Anthropic serve that work profitably.
Anthropic’s second-quarter improvement is visible in a straightforward comparison. Revenue increased from about $4.73 billion in the first quarter of 2026 to $11.56 billion in the second. Total operating expenses rose from about $6.40 billion to $10.58 billion. Quarter over quarter, revenue grew approximately 144 percent, while total operating expenses rose about 65 percent.
Figure 3. Total operating expenses include model development and service delivery, personnel, sales and administration. Rounded; unaudited. Source: S-1, p. 158.
The expense base was expanding rapidly. Revenue increased by enough to cover that larger base and turn a $1.667 billion operating loss into $974 million of operating profit. The first half as a whole still recorded an operating loss, but the second quarter established that the business could generate revenue above its reported operating expenses.
Operating profit is what remains after those expenses, before interest, other non-operating gains and losses, and taxes. The comparison includes the costs of developing models, employing staff and distributing the service. It is therefore more informative about the overall operation than a margin that leaves research spending out. It also remains distinct from net profit or cash available to pay future bills.
The timing of computing costs helps explain why this improvement can happen quickly. Anthropic says these expenses are generally fixed in the near term once committed resources become available. Revenue can then rise faster than that cost base as paid demand grows. The reported figures show the resulting improvement; they do not isolate how much came from better resource utilization or any particular product.
A subsequent delivery of computing capacity can change the comparison again. Understanding that schedule is essential to judging what the profitable quarter says about the years ahead.
Anthropic typically arranges computing capacity 12 to 18 months before it becomes available. Here, capacity means access to the computing resources used to develop and run its models. The company has to make those arrangements while the eventual mix of customer tasks and research needs is still uncertain.
Figure 4. Timing of new computing capacity. Payment and cancellation terms vary by contract. Source: S-1, pp. 145, 159 and 170–172.
If capacity arrives before enough paid work, the company may incur higher expenses while waiting for demand to catch up. If demand arrives first, the shortage can restrict the work it can serve. Anthropic also says short-term computing resources generally cost more per unit, so filling a gap at short notice can put pressure on margins.
Usage can change much faster than this procurement schedule. The filing reports that total customer spending in the two weeks ended January 2, 2026 was about 20 percent below the preceding two weeks. It discusses this alongside holiday patterns and pauses in software development. Even customers that continue using Claude can vary their activity considerably across short periods.
Minimum customer spending commitments help with this uncertainty, but they need to be assessed alongside the computing agreements. A customer revenue floor and an infrastructure commitment may cover different periods and amounts. Contracting for future capacity also does not mean paying the entire bill at signing; delivery schedules, payment obligations and cancellation rights depend on the agreement.
The commercial challenge is to match capacity with demand closely enough that useful work supports the resources being delivered. Strong long-term demand can coexist with a difficult quarter if those schedules do not line up. That is why another period of fast revenue growth, by itself, would not settle the question of lasting profitability.
More than 80 percent of Anthropic’s third-quarter 2026 revenue came from model versions released between April and September of that year. Recent releases are already central to what customers buy. Continued development therefore matters to today’s business, as well as future growth.
Set that product pace beside the 12-to-18-month capacity schedule. Resources arranged today may need to support a different mix of models and customer tasks by the time they arrive. Anthropic must plan beyond the workloads it can currently observe. A forecast of growing AI demand still leaves decisions about which resources to secure, when to bring them into use, and how much to devote to research.
Research spending belongs in this calculation because improved models can make additional customer work feasible. Cutting it could improve a near-term result while weakening future demand. Additional research earns its keep when it helps win or retain paid work, or lowers the cost of serving that work. Newer models’ popularity alone does not establish the return on developing them.
The second-quarter result matters because revenue covered the full reported operating cost base, including research, while Anthropic was still developing its products. The business earned an operating profit during rapid expansion. Whether it can keep doing so depends on how revenue responds as new models and new computing resources enter service.
The most useful follow-up will combine customer spending with those costs. Continued expansion among established customers would show that the opportunity extends beyond initial trials. Repeated operating profits through capacity deliveries would show that Anthropic can support that demand without allowing costs to outrun revenue. Neither requires the exceptional ninefold customer-spending ratio to continue indefinitely.
Anthropic can earn more from the same organization as Claude makes more work worth doing. To serve that expansion, it must commit to resources before the eventual workload is clear. Its next test is whether improvements in the product turn into paid usage soon enough to support the resources secured for them.
Based on Anthropic’s Form S-1 filed October 26, 2026, reviewed October 29: pp. 3–4 and 136 (retention); 139 and 143–145 (revenue mix, customers and usage); 134–135 and 158–159 (capacity and results); 170–172 (contracts); and 196 (Opus 4.5). Quarterly financials are unaudited; figures are rounded; Figure 2 is hypothetical. The offering remains preliminary. The Rakuten example comes from its June 26, 2025 case study and reports customer observations, not independently verified causal effects.
Inside the SpaceX S-1: A Comprehensive Analysis of the Largest IPO on Record
What Insurance Does a Public Company Actually Need?
