.png?width=500&height=250&name=%E7%9F%A5%E7%9A%84%E8%B2%A1%E7%94%A3%E4%BA%8B%E5%8B%99%E6%89%80%E3%82%A8%E3%83%9C%E3%83%AA%E3%82%AF%E3%82%B9%20(500%20%C3%97%20250%20px).png)
The era in which AI agents “use” computers has arrived. Leading the way is Anthropic’s “Computer...
Anthropic, known for Claude, is now one of the world’s leading AI companies. You might think that a company with such innovative technology would have filed a large number of patents—but upon closer inspection, a surprising fact emerges.In this article, a patent attorney well-versed in AI intellectual property will analyze Anthropic’s “patent strategy” by citing actual registered patents and their claims.
Rather than simply introducing the patents, this analysis takes a strategic perspective on “how leading AI companies manage their intellectual property.” It offers valuable insights for executives and IP managers grappling with the decision of whether to protect their company’s AI and software technologies through patents, academic papers, or trade secrets.
💡 Key Point: This article is part of the AI Agent Patent Series. For the basics, see the “Fundamentals” section; for a comparison of cases in Japan, the U.S., and Europe, see the “Case Studies” section; and for a detailed explanation of Salesforce’s specific patents, click here.
Table of Contents
- Conclusion | Anthropic Is an AI Company That “Does Not File Patents”
- Surprising Fact | Anthropic Has Surprisingly Few Patents
- Caution: The Trap of the “39” and “61” Figures
- Patents Actually Held by Anthropic | All Are “Computer-Operated Agents”
- Analyzing the Core Patent | US 12,430,150 B1 (The Heart of “Computer Use”)
- The Broadest Scope of Rights | US 12,566,913 B2
- Why Not Patent Flagship Research? | The “Publish, Don’t Patent” Strategy
- Patent Network Acquired Through M&A | The Significance of the Adept Acquisition
- Warning from Patent Attorneys | Just Because It Says “Claude” Doesn’t Mean It’s Covered by Anthropic’s Patents
- Four Lessons Japanese Companies Should Learn
- Frequently Asked Questions (FAQ)
Conclusion | Anthropic Is an AI Company That “Does Not File Patents”
● Research Conclusion: Among major AI companies, Anthropic files very few patent applications.
● Patents Held: All verifiable patents belong to a single technology family titled “computer-use agents” (derived from Adept, acquired in 2024; priority date: March 20, 2024).
● Technologies Not Patented: Flagship research such as Constitutional AI, RLHF/RLAIF, interpretability, and prompt caching has been published as papers rather than patented.
● The Essence of the Strategy: An extremely deliberate IP management approach: “Publish core research; patent product features defensively.”
● Patents Held: All verifiable patents belong to a single technology family titled “computer-use agents” (derived from Adept, acquired in 2024; priority date: March 20, 2024).
● Technologies Not Patented: Flagship research such as Constitutional AI, RLHF/RLAIF, interpretability, and prompt caching has been published as papers rather than patented.
● The Essence of the Strategy: An extremely deliberate IP management approach: “Publish core research; patent product features defensively.”
Surprising Fact | Anthropic Has Surprisingly Few Patents
A careful review of patent databases reveals that the number of U.S. patents attributed to Anthropic PBC (Anthropic’s corporate name) is limited to approximately six granted patents and about seven pending published applications (as of 2026; subject to verification). Furthermore, all of these relate to the same technical theme.
This is a strikingly small number compared to OpenAI—which is reported to have built a patent portfolio in the hundreds—or Google and Microsoft, which boast application numbers on an entirely different scale. Why does a world-class AI company file so few patents?—Herein lies the strategy.
| Perspective | Anthropic’s Trends | Major Tech Companies (General Overview) |
|---|---|---|
| Number of Applications | Very low (around a dozen) | High (several hundred to several thousand) |
| Target Technologies | Focused on computer-operated agents | Covers a wide range of technical fields |
| Treatment of Core Research | Published as papers (not patented) | Often Patented |
| Acquisition History | Acquisitions (Adept) are the primary source | Primarily through the company’s own systematic patent filings |
Caution: The Pitfall of the “39” and “61” Figures
Beware of miscounts: Some patent analysis services report Anthropic’s patents as “39” or “61,” but these figures are inflated due to confusion with other companies (errors in name matching).Specifically, patents held by the UK-based company “Anthropics Technology Limited” (which develops the photo-editing software “PortraitPro”—note: “Anthropics,” not “Anthropic”) and unrelated machine learning patents have been mistakenly included in the tally.
This is a fundamental and important lesson in patent research. If you confuse patent holders simply because their company names are similar, your research results will be fundamentally skewed. Anthropic’s true patent portfolio is far smaller than the inflated figures suggest and has a clear focus.
Patents Actually Held by Anthropic | All Related to “Computer-Controlled Agents”
Without exception, Anthropic’s verifiable patents relate to technology where “AI agents automatically operate a multimodal (image + text) user interface.” This corresponds exactly to “Computer Use” (the feature where Claude operates a PC using a mouse and keyboard), which Anthropic announced in October 2024.
The following are representative verified patents and patent applications (applicant: Anthropic PBC; priority date: March 20, 2024, for all).
| Number | Title (Summary Translation) | Classification |
|---|---|---|
| US 12,430,150 B1 | Runtime Architecture Interacting with an Agent to Automate Multimodal Interface Workflows | Registered Patent (September 30, 2025) |
| US 12,566,913 B2 | AI Agent for Automating Multimodal Interface Tasks | Registered Patent (March 3, 2026) |
| US 2025/0299023 A1 | System and Method for Configuring an AI Agent to Automate Multimodal Interface Workflows | Published Application |
| US 2025/0299074 A1 | Data Flow Logic for Providing an AI Agent That Automates Multimodal Software Usage | Published Application |
💡 Key Point: Anthropic’s patent portfolio consists of “defensive, function-specific” patents that correspond one-to-one with its products (Computer Use). Rather than casting a wide net, the company strategically acquires patents to protect its flagship features.
Reading the Core Patent | US 12,430,150 B1 (The Heart of “Computer Use”)
The registered patent US 12,430,150 B1, titled “Runtime architecture for interfacing with agents to automate multimodal interface workflows,” can be considered Anthropic’s foundational patent. Below is a direct quote of Claim 1 (independent claim) from the original text.
Key Points of This Claim
Noteworthy is the division of processing between the “client-side” and “server-side,” connected by a multi-stage transformation: “agent specification → intermediate representation → agent call → execution function → execution command.”
Rather than simply presenting the idea that “AI operates a PC,” the patent describes a concrete software architecture that includes a division of roles between client and server, an abstraction layer in the form of an intermediate representation, and the translation of functions into commands. This level of specificity is key to meeting U.S. patent eligibility requirements (Alice/Mayo) and the requirements for software-related inventions in Japan.
Broadest Scope of Protection | US 12,566,913 B2
The registered patent US 12,566,913 B2, which belongs to the same patent family, addresses “interface automation” at a higher conceptual level. The following is a quote from its abstract.
This description succinctly captures the essence of computer-use technology: “a machine replicates user-actuated actions on an interface.” This is considered one of Anthropic’s broadest claims in this field.
Why Not Patent Flagship Research? | The “Publish, Don’t Patent” Strategy
This brings us back to the biggest mystery. Anthropic has developed numerous industry-leading technologies, such as Constitutional AI, RLHF/RLAIF, interpretability, prompt caching, and batch processing. Despite this, the company does not patent these technologies but instead publishes them as research papers and technical documents.
This is not because they “couldn’t obtain patents,” but rather is believed to be an intentional strategy. The rationale behind this is as follows.
| If They Were to Patent These Technologies | If Published as a Paper |
|---|---|
| Gain a 20-year exclusive right | No exclusive rights are obtained |
| Obligation to disclose the technology in detail in the specification | You can disclose the technology to the extent you choose |
| Can enforce rights and generate licensing revenue | Gain a head start in branding, talent acquisition, and standardization |
| Legally deter competitors from following suit | Promoting the development of the entire ecosystem |
| Capture insights on AI safety | Broadly sharing safety insights (in line with our philosophy) |
💡 Key Point: For Anthropic, which champions AI safety as a core principle, widely sharing safety research through academic papers aligns better with both its philosophy and brand than hoarding it behind patents. On the other hand, the company defensively patents features that differentiate its products (Computer Use).This clear distinction between “what to disclose and what to keep proprietary” is the core of Anthropic’s IP strategy.
Patent Portfolio Acquired Through M&A | The Significance of the Adept Acquisition
Another key strategic point is that the majority of Anthropic’s patents were acquired through mergers and acquisitions (M&A), rather than through the company’s own systematic filing efforts.
In 2024, Anthropic acquired key members and technology from the AI agent company Adept (an “acqui-hire”).The list of inventors for the aforementioned patent portfolio includes David Luan, Erich Elsen, and Curtis Hawthorne, all former Adept employees. In other words, Anthropic acquired both the “technology that enables Computer Use” and the “patents that protect it” in a single transaction through the acquisition.
Implications for IP Strategy: A patent portfolio can be built not only by filing applications in-house from scratch but also by acquiring companies that possess the technology necessary for one’s business.The Anthropic case is an excellent example of a speed-focused IP strategy in which “product launch (Computer Use, October 2024)” and “acquisition of the underlying technology and patents (Adept, 2024)” were closely linked.
Warning from a Patent Attorney: Just Because It Says “Claude” Doesn’t Mean It’s an Anthropic Patent
There is a critically important point to note in practice. Even if a patent specification mentions “Claude,” “Anthropic,” or “large language model,” the patent holder (applicant) is not necessarily Anthropic.
Beware of misidentifying the rights holder: During our research, we found numerous patents that mentioned “Claude” or similar terms but were actually owned by other companies.Examples: U.S. Patent US 12,353,469 B1, concerning the verification and citation of language model outputs, is an Amazon patent; U.S. Patent Application US 2025/0278287, concerning UI automation, is filed by Workday;and the application WO 2025/188958 regarding LMM/LLM alignment is filed by Seekr Technologies.
To accurately identify the patent owner, it is essential to check the “assignee” field for each entry individually. This is the absolute foundation for avoiding errors in judgment during FTO (Freedom to Operate) searches and competitive analysis. The risk of misidentification increases for patents containing trending keywords, such as “AI.”
Four Lessons Japanese Companies Should Learn
① Consciously distinguish between “patents vs. research papers and know-how.” Patenting everything is not the right approach. Core functions you wish to monopolize should be patented, while research where sharing is advantageous and know-how that benefits from confidentiality should be protected by other means. Categorizing based on business strategy is crucial.
② Defensive and function-specific applications are also effective. A high-volume filing strategy is not the only intellectual property strategy. A “small but elite” approach—precisely protecting your company’s differentiating features—is also a sound strategy.
③ Consider acquiring patents through mergers and acquisitions. If you cannot devote time to filing your own applications, one option is to build your portfolio quickly by acquiring a company that possesses both the technology and the patents (acqui-hire).
④ Always verify the rights holder during patent searches. Mistakes caused by similar company names can be fatal. Be sure to thoroughly verify the assignee for each case.
Should your company’s technology be patented? Or should it remain as a research paper or know-how?
Patent attorneys with expertise in the IT, software, and AI fields provide comprehensive support—from assessing patentability to designing “patent vs. confidentiality” strategies, filing applications, and conducting FTO searches. We offer a free initial consultation.
Schedule a Free Initial Consultation IT & AI Intellectual Property ServicesFrequently Asked Questions (FAQ)
Q. Does Anthropic hold many patents?
A. No. Among major AI companies, Anthropic is a “light filer” with very few patent applications. The number of verifiable U.S. patents—including both granted and pending applications—is only in the low teens, and all of them belong to a single technology family: “computer-use agents.”This stands in contrast to reports that OpenAI, Google, and Microsoft have filed in the hundreds.
Q. What kind of technology do Anthropic’s patents cover?
A. All of Anthropic’s verifiable patents pertain to technology where “an AI agent automatically operates a multimodal (image + text) user interface.” This corresponds to the “Computer Use” feature (which allows Claude to operate a computer) that Anthropic announced in 2024.These patents were filed by the team from Adept, a company acquired by Anthropic in 2024, and all have a priority date of March 20, 2024.
Q. Why doesn’t Anthropic patent technologies like Constitutional AI?
A. Anthropic tends to publish its flagship research—such as Constitutional AI, RLHF/RLAIF, interpretability, and prompt caching—as papers and technical documents rather than patents.This is seen as a reflection of the company’s philosophy of widely sharing insights on AI safety and its commitment to open development over monopolization through patenting (the “publish, don’t patent” strategy).
Q. Are all patents labeled “Claude” or “Anthropic” owned by Anthropic?
A. No. Even if the specification mentions “Claude” or “Anthropic,” the assignee is often a different company. For example, the U.S. patent (US 12,353,469 B1) regarding the verification and citation of language model outputs is owned by Amazon.To determine the patent holder, you must check the “assignee” field for each patent individually.
Q. Is a low number of patent applications a disadvantage for a company?
A. It’s not possible to generalize. While patents offer the benefit of “competitive advantage through exclusivity,” filing a patent application comes at the cost of disclosing the technology in the specification. A strategy of publishing research in academic papers to gain a lead in brand building, talent acquisition, and standardization is also effective.The key is to consciously distinguish between “technologies to be patented” and “technologies to be disclosed as know-how or in academic papers” in accordance with your company’s business strategy.
Important Note Regarding This Article: This article is a general explanation based on publicly available patent information as of 2026. Patent numbers, counts, and statuses are subject to change over time, and the scope of rights is determined by the claims and historical data for each patent.When making business decisions (such as FTO analysis, infringement analysis, or filing strategies), be sure to consult the latest published patent applications and prosecution history, as well as seek individual consultation with experts. While the cited claims and abstracts are based on published patent application data, we recommend verifying them against the official USPTO records for legally significant purposes.
Recommended Reading (AI Agent Patent Series): • Can
AI Agent Technology Be Patented? Examination Standards and Key Points for Granting Patents (Basics) • Can
AI Agents Be Patented? Patent Cases and Examination Practices in Japan, the U.S., and Europe (Case Studies) • Decoding
Salesforce’s Multi-Agent Patents (Individual Analysis)
AI Agent Technology Be Patented? Examination Standards and Key Points for Granting Patents (Basics) • Can
AI Agents Be Patented? Patent Cases and Examination Practices in Japan, the U.S., and Europe (Case Studies) • Decoding
Salesforce’s Multi-Agent Patents (Individual Analysis)
Source
- US 12,430,150 B1 (Google Patents): https://patents.google.com/patent/US12430150B1/en
- US 12,566,913 B2 (Google Patents): https://patents.google.com/patent/US12566913B2/en
- List of Patents Assigned to Anthropic, PBC (Justia): https://patents.justia.com/assignee/anthropic-pbc
