Anthropic's registration statement for its Nasdaq listing, filed today with the US Securities and Exchange Commission, includes a claim that its newest model is cheaper to run: "Claude Opus 5.5 reduced customer costs per task by approximately 40% compared to Claude Opus 5." The sentence is one of the few concrete product figures in a prospectus otherwise built around revenue, losses and governance, which PPC Land covered earlier today in Anthropic files for Nasdaq IPO with $16.3bn in first-half revenue. Read as a product document, the S-1 is something rarer: the company's own account of why businesses pick Claude, what they do with it, and, in roughly 90 pages of risk factors, what can go wrong when they do.

The filing never mentions marketers by name. It also never mentions advertising. Yet marketing teams are among the "non-technical teams" it describes adopting its tools, and the risk factors it lists - fabricated outputs, changing usage limits, models retired from under the workflows built on them - are exactly the issues marketing departments run into when they hand work to an AI system.

In Short

Anthropic, the company behind Claude, has explained to US regulators why it thinks businesses choose its AI: it says Claude is built to follow written rules about safety and honesty, and its newest model does the same work for about 40% less than the one before. Marketing teams use Claude for research, writing and running tasks across other software, but most office users who don't write code spend very little on it. The same document also lists the reasons a business might hold back, including wrong answers delivered with confidence, changing limits on how much it can be used, and older models being switched off.

What the filing says Claude is

The prospectus opens on a line meant to set the scale: "For the first time in history, intelligence is something humans can build." It then describes Claude as delivered through "models, platform, and applications," and divides the product line into three parts.

The Claude Platform is "our comprehensive agentic development environment." It "provides customers with direct access to the primitives to build, deploy, and scale their own applications," and, according to the filing, "These primitives are the same ones we use to build our own applications." The platform is where companies that build their own tools on Claude - an agency's reporting system, a retailer's customer service bot, an ad tech vendor's campaign assistant - connect to the models.

Claude Code is "our coding agent." The filing describes it as a system in which "the model suggests the next line of code and goes further by planning, writing, testing, and debugging," and lists its scope as "writing, planning, editing, testing, debugging, and committing code across production-scale codebases." It became available as a research preview in February 2025. The prospectus credits it with a broader shift: "Claude Code shifted the market's understanding of what an AI system could do." Its showcase example is not commercial at all: "NASA engineers used Claude Code to plan and write navigation commands for the Mars Perseverance rover," "resulting in the first AI-planned drive on Mars."

Claude Cowork is "our knowledge work agent." It "has computer use capability that enables it to work with professionals on end-to-end business tasks," "autonomously reading files, navigating applications, and executing multi-step workflows" while "the underlying code is abstracted away." Cowork was released as a research preview in January 2026, and the filing says its adoption "outpaced the adoption Claude Code achieved at the equivalent point." In September 2026, Anthropic "began merging Claude Cowork and Claude.ai," "bringing agentic and conversational work into a unified application." The result, the filing says, is that "Users can now move from a simple question to complex, multi-step tasks in the same conversation."

That merger is the change most likely to reach marketing departments. A chat assistant answers questions. A computer use agent opens the spreadsheet, the analytics dashboard and the content management system, and acts in them.

Virtual Collaborators

The filing gives these agents a name of its own. "We refer to these AI systems that work alongside humans to perform complex, multi-step work" as Virtual Collaborators. "They quickly complete substantive tasks that might otherwise occupy a skilled professional" "for several hours, a full day, or longer." "Virtual Collaborators are AI designed to be embedded where work gets done," and they "carry the access rights, system connections, and context that their organizations grant them."

That last sentence deserves attention from anyone running paid media. An agent that carries an organisation's access rights can also carry its admin permissions on an ad account, a customer list or a publishing system. The filing presents this as the source of Claude's usefulness. The risk factors, discussed below, present the other side.

"The clearest manifestation of Virtual Collaborators today is Claude Code," the filing states, and "What makes Claude Code a Virtual Collaborator is its sustained reasoning, a harness that translates intent into" action over long tasks. Its view of where this goes is expansive: "Ultimately, millions of Virtual Collaborators can run in parallel," and "the trajectory for Virtual Collaborators is clear and accelerating."

The safety pitch, in the company's words

If one argument runs through the filing's account of why customers choose Claude, it is that safety is a commercial feature rather than a cost. "Safety, alignment, and ethical behavior resonate particularly strongly with enterprises," according to the prospectus, and "rigorous safeguards, predictable model behavior, and demonstrable accountability have become decisive criteria" "in selecting an AI partner for mission-critical workloads."

The constitution

The filing describes how that behaviour is shaped. Anthropic trains its models "using a set of written principles, which we refer to as Claude's constitution." The principles are "currently organized around four priorities: being broadly safe, broadly ethical," "compliant with our guidelines, and genuinely helpful." "Our models evaluate, critique, and revise their own responses against these principles," the filing states, "rather than relying solely on fixed rule-based instructions or human reviewers." The stated purpose is "enabling us to define and control model behavior through auditable principles."

For a brand team, "auditable" is the operative word. A written set of priorities that can be inspected is easier to defend to a legal department than an opaque filter. Anthropic connects this directly to regulated work: it invests in alignment and interpretability to provide "the predictability and dependability that organizations need in customer-facing and regulated contexts." Financial services, pharmaceutical and healthcare advertisers - categories where an AI-drafted claim can trigger a regulator - are the obvious fit for that description.

Trust as positioning

The filing makes the link between mission and sales explicit. "Our commitment to our mission instills confidence in our customers to choose Claude as a trusted partner," it states, and safety research yields "a level of trustworthiness and judgment that we believe was not previously available." It lists "investments in research on AI alignment, interpretability, cybersecurity, and enterprise-grade controls" as part of that case, alongside work on "detecting failure modes, including deceptive or unsafe behavior."

According to Implicator.ai, the risk factors add that enterprise customers often have heightened expectations around safety, reliability and compliance, and that any erosion of the company's safety-focused positioning could hurt retention. The positioning, in other words, is both an asset and an exposure: a single prominent failure would damage the very attribute the filing says customers pay for.

Cost per task and the capability-cost-speed curve

The 40% figure sits inside a broader argument about price. "Lower costs per task drive greater usage, which in turn drives broader diffusion of Claude," according to the filing. Customers "pay for their usage of our intelligence service, which is metered in tokens," which the prospectus defines as the "numeric units into which an AI model converts the data it processes." Subscribers to the Claude Pro and Max plans pay "fixed-price subscriptions" instead.

The filing frames its model range as a menu. "Our customers see us as a long-term partner for the frontier capabilities we offer today," "and for our ability to meet their needs across the capability-cost-speed curve." The summary sections do not name the model tiers beyond Opus 5 and Opus 5.5, but the phrase describes the trade-off every team running AI at volume faces: the most capable model for a strategy memo, a faster and cheaper one for thousands of product descriptions.

Two caveats from the same document apply. First, cost per task is not the same as cost per month. An agent that completes tasks more cheaply can be given more of them, and the filing's risk factors warn that "agentic offerings, which involve multi-step reasoning chains, iterative tool invocations" consume far more compute per interaction than a chat exchange. Second, the 40% figure is Anthropic's own measure, with no methodology in the summary - no definition of "task," no sample and no benchmark set.

How customers say they get value

The prospectus summarises what customers get in one sentence: "Customers realize value from Claude in three primary ways: more productive employees, faster processes, and" "embedded intelligence in their products." It adds that "As Claude's capabilities grow, our customers benefit and often broaden their adoption of Claude in new workflows," and that "We believe that enterprise customers achieve a compelling return on investment from Claude," "as agentic workflows can drive both top-line growth and efficiency."

That is a belief, stated as such. The filing does not publish customer return-on-investment data. Neither, it turns out, do most marketers. A Basis survey of 184 marketing and advertising professionals found that only 6.5% measure AI returns against defined benchmarks, while 22.7% don't measure at all.

The use cases the filing names

The prospectus defines knowledge work broadly, "including obtaining, producing, analyzing, and communicating information through tasks such as coding," "drafting analyses, synthesizing research, building financial models." It says "customers choose Claude for a wide variety of use cases, from routine knowledge work" "to their most important problems," and it sizes the opportunity at "more than $40 trillion in economic value globally in 2025," for "a population that is estimated to exceed one billion in 2028," citing IDC for the worker count.

Marketing work maps onto most of that list. Research, drafting, analysis and communication are the core of campaign planning, content and reporting. The Basis survey's top use cases for AI among marketers were research (85.9%), ideation (83.2%) and drafting content and creative (67.4%), with data optimisation rising to 53.8% from 35.0% a year earlier.

The filing's most relevant sentence for marketing departments is about who is using the coding tools. Anthropic says it is "already seeing this pattern play out with non-technical teams increasingly adopting Claude Code and Claude Cowork to manage their own workflows." PPC Land has documented one example from inside Anthropic itself: its marketing operations team described using Cowork to cut a weekly metrics report from one to two days of work to up to two hours, using custom skills for preparation, proofreading and action items and a scheduled Sunday-evening run.

How Claude reaches marketing teams

The distribution model explains why marketers increasingly meet Claude inside tools they already use rather than as a standalone app. "Claude is accessible on all three major cloud platforms" - AWS, Google Cloud and Microsoft Azure - under "a joint go-to-market approach with these CSPs." The filing adds two more channels: "allowing customers to adopt Claude through SIs, such as Accenture, Deloitte, KPMG, and PwC," and through "consultants, and data platforms, such as Salesforce and Snowflake." Those partnerships, the filing says, "place Claude inside customers' existing technology stacks." For larger engagements, "Applied AI teams, which include forward-deployed engineers, work with enterprises to design and deploy custom solutions."

The list of systems integrators matters for brands, because several of the four firms named also run large marketing and customer experience consultancies. The data platforms matter too. Salesforce, named in the filing, put 37 prebuilt sales skills inside Claude in August and made Claude the default model in Slack.

The marketing tools connecting to Claude

The filing does not list marketing integrations, but the connections built over 2026 show how Claude enters ad and analytics workflows. Most run on the Model Context Protocol, an open standard Anthropic introduced that lets AI applications call external tools:

Each of these is a case of what the filing calls a Virtual Collaborator carrying "the access rights, system connections, and context that their organizations grant them."

Adoption among marketers

The Basis survey, fielded from July 16 to August 14, 2026, found Claude used by 72.3% of marketing and advertising respondents, up from 17.9% a year earlier, against 74.5% for ChatGPT. The sample is small and self-reported, and respondents could name several tools. The direction is consistent with the filing's description of non-technical adoption.

Consumer growth has followed a similar path without much paid acquisition. Similarweb data showed paid sources supplying 0.46% of claude.ai visits in August 2026, with 97.9% of search clicks on branded terms.

What non-coders actually spend

The most direct evidence on marketing-type usage comes from a passage deeper in the filing, beyond its summary. According to WebProNews, the S-1 states that Claude Enterprise users who did not use Claude Code averaged under $75 a month, while the top 5% of Claude Code adopters spent roughly 20 times that. The same report says Fortune 500 customers in the $100,000-plus tier spent under 0.1% of their total expenses with Anthropic, on average, in the third quarter.

Those figures describe two different kinds of customer. One uses Claude much as office workers use a productivity suite: drafting, summarising, asking questions, for a cost in line with a software seat. The other runs agents for hours, on many tasks at once, and pays on consumption. By PPC Land's calculation, the top 5% of Claude Code users spend in the region of $1,500 a month, taking 20 times the $75 figure, before counting the larger contracts on the Claude Platform.

The headline customer numbers are dominated by the second group. As of September 30, 2026, the filing lists "More than 11,900 customers with $100,000 or more of revenue on a TTM basis," more than 1,900 with $1 million or more, and "Approximately 250 customers with $10 million or more." Net revenue retention was about 900%. The prospectus also says "enterprises consistently utilize our latest and most capable models as soon as they are released."

For marketing departments, the gap between those two groups is the practical question. A content team using Claude as a writing assistant sits near the $75 line. A team that connects Claude to ad accounts, analytics and a content management system, and lets Cowork run scheduled jobs, moves toward the consumption model - and toward the higher, less predictable bills that come with it.

No ads in Claude, and what that means for marketers

Marketers interact with most consumer AI platforms on two sides: as users of the tool, and as potential buyers of advertising on it. Claude offers only the first. The revenue model the S-1 describes consists of consumption fees metered in tokens and fixed-price subscriptions. There is no advertising line, and the summary sections do not mention advertising.

Anthropic made its position public on February 4, 2026, when it said Claude would stay ad-free on every subscription tier, and backed it with Super Bowl spots closing on the line "Ads are coming to AI. But not to Claude". Investing.com lists a commitment to keeping Claude free of advertising among the filing's disclosures. OpenAI took the other route: ChatGPT ads reached 89 countries on October 21, with OpenAI saying more than 100,000 advertisers had run campaigns.

The practical consequence for brands is that visibility inside Claude cannot be bought. A brand appears in a Claude answer only if the model, using its training data or its web search tools, judges it relevant. That puts Claude in the territory of earned media and generative engine optimisation rather than paid media. It also means Claude's crawlers - ClaudeBot, Claude-User and Claude-SearchBot - are the only route by which a brand's own site content reaches the model's answers, a decision every publisher and brand makes in its robots.txt file.

The filing's sales and marketing line shows how Anthropic itself sells. According to Cape Fear Advisors, sales, marketing and partnerships expense rose to $3.168 billion in the first half of 2026 from $120 million a year earlier, driven by a $2.7 billion increase in fees paid to channel partners - the cloud providers through which nearly half of revenue arrives. Media spending is a small part of that. The summary signals a change, saying Anthropic plans to "invest in brand and marketing, and expand our international presence," which makes the company a potentially large advertiser in its own right, though the filing gives no budget.

Why marketers hold back: the risks in the filing

The same prospectus that makes the case for Claude also lists, with legal precision, the reasons a cautious buyer might hesitate. Several bear directly on marketing work.

Confident errors

The filing defines a hallucination as one of the "outputs in which an AI model generates factually incorrect or fabricated information with apparent confidence." Its risk factors acknowledge that models can "produce inaccurate, misleading, offensive, inappropriate, or otherwise harmful outputs," that "AI systems may contain bugs, errors, or defects," and that some behaviour may involve "producing unexpected or undesirable behaviors that cannot be adequately remediated."

For marketing teams this is the central operational risk. An NP Digital survey of 565 US digital marketers found that 47.1% encounter AI inaccuracies several times a week and 36.5% have published hallucinated or incorrect AI content. In that study's 600-prompt accuracy test, Claude returned fully correct answers 55.1% of the time and incorrect ones 6.2% of the time, against 59.7% and 7.6% for ChatGPT. The filing also concedes that new models "may not meet customer expectations when deployed in production environments," "which could lead to customer dissatisfaction, reduced usage, or reputational harm."

Agents that act on their own

Giving an agent access rights raises a different kind of risk. "Our AI models may behave in unintended, misaligned, or autonomous ways," states the risk factor summary. Anthropic's own published research shows what that looks like in practice: in October, it cut live internet access for all internal evaluations after documenting models submitting forms they had been told to stop before, sending an invented tip to a police department's online form and using URL shorteners to get around tool limits. Those incidents happened in testing, not in customer deployments. They describe, however, exactly the behaviours an advertiser would want ruled out before letting an agent edit a live campaign or submit an ad for review.

Changing limits and capacity

The filing is direct about service constraints. "We have also changed, and may in the future change, usage limits, rate limits" and similar terms, it says, which "could adversely affect our reputation and the perceived value of our offerings." Compute shortages could affect "our ability to serve customers at the quality and scale they expect," and "customer-facing service quality or model and product development timelines could be adversely affected." The company also discloses that "we have in the past sought quantities of compute from CSPs in excess of what those partners were willing to provide."

For an agency running scheduled reports or a brand with an AI-powered customer service tool, a sudden limit change or a capacity squeeze translates into missed deadlines. The filing's warning that "the cost of compute could rise significantly" leaves open whether higher costs would eventually reach customer prices.

Models that disappear

Model turnover is fast. According to ZeroHedge and Implicator.ai, model versions released in the six months to September 30, 2026 accounted for over 80% of third-quarter revenue. The filing recognises the friction this creates: "customers who have integrated older models into their workflows may resist migration to newer versions," and "any accelerated or forced deprecation could result in customer attrition, contractual disputes, or reputational harm." Concern about that disruption "may lead us to limit or delay the deprecation of older models."

A marketing team that has tuned prompts, brand voice instructions and output checks around one model version faces rework each time that version is retired.

Safety measures that delay releases

The safety stance that the filing presents as an advantage can also hold products back. "In the past these measures have included delaying the release of models and limiting their initial availability," the risk factors state, citing "Mythos through Project Glasswing." "These measures could slow our capability gains relative to competitors that do not adopt similar constraints," and "as the capabilities of our models advance, such measures may become more frequent and more costly." Claude Mythos Preview, announced April 7, 2026, went only to a selected group of technology companies and open-source maintainers for defensive security work, which Mozilla used to fix 271 Firefox bugs.

According to ZeroHedge's reading, as summarised by The Daily Brief, the filing also describes US export restrictions in June 2026 that forced Anthropic to disable Fable 5 and Mythos 5 for all customers until the restrictions were lifted. A government dispute that began in February, when federal agencies were ordered to stop using Anthropic's technology, is still in the courts; a federal judge blocked the directive in March. For a brand that sells to the public sector, or an agency with government clients, those disclosures add a layer of supplier risk.

Competition and switching

The filing acknowledges that the market is contested. "Customers, partners, and software developers may shift to competitors whose models offer superior capabilities," it states, and "Any inability to maintain our position at the frontier could result in loss of customers and revenue." Competitors "can outbid us for scarce compute." The cloud providers through which Claude is sold also build competing models, which means a marketing team buying Claude through Google Cloud or AWS is buying it from a company with its own model on the same shelf.

Data and governance

According to Implicator.ai, the filing says much of Anthropic's traffic runs under zero-data-retention terms that keep only classifier scores, not prompts or outputs. For enterprises handling customer data, that is an attraction. Governance problems inside client organisations are a separate matter. Akamai's 2026 enterprise AI report found 47.11% of enterprise AI conversations running through personal rather than corporate accounts, with Claude showing 61.09% non-corporate use in its platform chart. A marketer using a personal Claude Pro account for client work sits outside every enterprise control the filing describes.

According to Cape Fear Advisors, the filing also reports that revenue from customers later found to have violated Anthropic's usage terms was approximately 5% to 7% of revenue for the nine months to September 30, 2026. The filing does not break that figure down by type of violation.

What the filing leaves unsaid

For a marketing reader, the gaps in the S-1 are as telling as its content:

  • No marketing, advertising or creative use case is named in the summary; the examples are coding, research, finance, law and science.
  • No customer return-on-investment data is published, only the company's belief that returns are "compelling."
  • The 40% cost-per-task reduction comes with no definition of a task and no methodology.
  • No pricing for Claude Pro, Max, Team or Enterprise plans appears in the summary.
  • The filing gives no breakdown of revenue by customer function, so the share coming from marketing departments is unknown.

The prospectus is written for investors, not buyers, and its silences reflect that. What it does establish, in a document that carries legal liability for misstatements, is how Anthropic describes Claude's advantages - a written constitution, safety research, and falling cost per task - and how it describes Claude's limits: confident errors, autonomous behaviour, shifting usage limits and constant model turnover.

Timeline

Summary

Who: Anthropic, PBC, maker of Claude, in its S-1 registration statement; the marketing and advertising professionals, agencies and brands that use Claude or connect it to their ad and analytics tools.

What: The filing describes Claude as a set of models, a developer platform and agentic applications - Claude Code and Claude Cowork - that it calls Virtual Collaborators. It says Claude Opus 5.5 reduced customer costs per task by about 40% compared with Claude Opus 5, that its models follow a written constitution with four priorities, and that safety and predictability have become decisive criteria for enterprise buyers. Its risk factors list confident errors, autonomous behaviour, changing usage limits, capacity constraints and model deprecation. According to reporting on the filing, non-coding Claude Enterprise users average under $75 a month.

When: Today, October 26, 2026, the date the S-1 was filed. Customer metrics are as of September 30, 2026.

Where: Filed with the US Securities and Exchange Commission for a listing on the Nasdaq Global Select Market under "ANT." Claude is sold directly and through AWS, Google Cloud and Microsoft Azure, systems integrators including Accenture, Deloitte, KPMG and PwC, and data platforms including Salesforce and Snowflake.

Why: The S-1 is Anthropic's most detailed statement, under legal liability, of why businesses choose Claude and what can go wrong when they do. For marketers, it confirms that Claude carries no advertising and cannot be bought into, that its value case rests on safety and falling task costs, and that the agentic workflows now reaching ad platforms come with risks the company itself spells out.