evorix blog

[In-Depth Explanation by a Patent Attorney] Analyzing OpenAI’s “Custom GPTs” Patent US 12,406,207 B2 | “Model Builder”: AI Creating AI

The ability to create a unique AI assistant with specialized knowledge without writing a single line of code—OpenAI’s “GPTs (Custom GPT)” has dramatically expanded the scope of AI adoption.So, how is this mechanism for “creating your own AI” technically implemented, and how is it protected by patents? The answer lies in the registered patent US 12,406,207 B2, “Systems and methods for generating customized AI models, which this article explores in depth.

The most notable aspect of this patent is the concept of a “Model Builder”—an AI designed to create other AIs (a meta-AI). Based on user requests, the AI itself assembles custom agents using structured procedures.This approach stands in contrast to Anthropic’s agent-building patent (where humans describe agents using a dedicated language), which we covered previously. A patent attorney specializing in AI intellectual property will explain this while citing the actual claims.

💡 Key Point: This article is part of our AI agent patent series. For a comparison of the “construction” layer, please see the article on Anthropic’s agent construction patent; for examination practices, please refer to the case studies from Japan, the U.S., and Europe.

Table of Contents

  1. 30-Second Summary | “AI Creates AI” Patents
  2. Basic Patent Information
  3. Background | Building Custom AI Is Complex and Costly
  4. Key Point 1 | Model Builder = “AI Expert That Creates AI”
  5. Key Point 2 | A Structured 4-Step Creation Process
  6. Key Point 3 | Fine-tuning + Instructions + Generation of a Dedicated Interface
  7. Key Point 4 | Safety Checks via the “Reviewer Model”
  8. Reading Independent Claim 1 Clause by Clause
  9. Two Methods for Creating Agents | “AI-Generated” vs. “Human-Written”
  10. How It Is Evaluated in Japanese, U.S., and European Examinations
  11. Lessons for Your Own Patent Applications
  12. Frequently Asked Questions (FAQ)

30-Second Summary | “AI Creates AI” Patent

● What the Patent Is About: A system that generates users’ own custom AI (agents). Compatible with OpenAI’s “GPTs.”
● Key Points: ① “Expert AI that creates AI” = Model Builder; ② A four-step process: establishing behavior → creating a profile → refining prompts; ③ Fine-tuning the base model + providing instructions + generating a dedicated interface; ④ Safety checks performed by a reviewer model.
● Patent Holder: OpenAI OpCo, LLC.
● Status: U.S. registered patent (registered September 2, 2025; 19 claims total; Track One expedited examination).

Basic Patent Information

Item Content
Patent Number US 12,406,207 B2
Title of the Invention Systems and Methods for Generating Customized AI Models
Registration Date September 2, 2025
Priority Date September 25, 2023
Applicant OpenAI OpCo, LLC
Inventors Nicholas Turley, Thomas Dimson, Olivier Godement, Michal Pokrass
Number of Claims 19 (Independent: Claims 1, 9, and 14)
Common Name GPTs (Custom GPT)
Status Registered Patent (Obtained through Track One Expedited Examination)

Background | Building Custom AI Is Complex and Costly

Large language models are powerful, but as the specification points out, they present the challenge of being “complex to master and requiring significant resources to operate.” In reality, building AI specialized for specific tasks from scratch has faced high barriers in terms of both expertise and cost.

This patent provides a mechanism for efficiently creating, evaluating, generating, and deploying “specialized custom models” based on provided expertise, capabilities, and instructions. The specification highlights technical advantages such as improved efficiency, reduced resource consumption, and enhanced connectivity.

Core Concept 1 | Model Builder = “AI that acts as an expert in building AI”

At the heart of this patent is the “model builder.” Claim 1 defines this as a “language model composed of a specific sequence of instructions for forming a custom model” that is instructed to “act as an expert in creating custom models.”

💡 Key Point: In other words, the Model Builder is “AI for creating AI (meta-AI).” When a user specifies, “I want an AI like this,” this meta-AI acts like an expert to assemble a custom agent.What makes this patent innovative is that it secures intellectual property rights over the nested structure of “AI generating AI” itself as a specific technical means.

Core Point 2 | A Structured 4-Step Creation Process

Model Builder’s “specific sequence of instructions” is explicitly outlined in Claim 1 as four steps. This is a major feature of the claim.

[Model Builder’s Command Sequence (Claim 1)] Step 1: Establish Behavior └ Define the behavior of the custom model based on features ▼ Step 2: Generate Profile └ Generate a profile for the custom model ▼ Step 3: Generate Refinement Prompt └ Generate a “refinement prompt” that specifies parameters for data extraction from the knowledge base, templates, expected output, and interface configuration ▼ + Instructions to execute each step “in order and without skipping”

💡 Key Point: It is important to note that the claims include the specific limitation to “execute each step in order and without skipping.”By defining the AI creation process as a strict procedure rather than “vague instructions,” the invention is made concrete, thereby avoiding rejection based on abstract ideas.

Core Element 3 | Fine-Tuning + Instructions + Dedicated Interface Generation

Based on the user’s characteristics (knowledge base + capabilities), the model builder performs the following three actions:

Operations Details
Fine-Tuning Fine-tunes the base model using the “knowledge base” to endow it with specialized knowledge
Instruction Generation Generate a set of commands that constitute a “capability”
Interface Generation Generate a dedicated interface for interacting with the custom model

This process generates a specialized custom agent and its dedicated UI from a general-purpose model in a single step. The specification document includes application examples such as a custom model for financial operations (connected to financial service APIs).

Key Point ④ | Safety Checks via “Reviewer Models”

The specification also mentions a mechanism in which “reviewer models” identify potentially harmful combinations within user settings and flag them for review.

Allowing anyone to create custom AI also creates a risk of misuse. This patent takes into account elements essential to the democratization of AI, such as balancing creative freedom with safety. The design, which incorporates safety mechanisms into the scope of the claims, serves as a practical reference.

A Clause-by-Clause Analysis of Independent Claim 1

US 12,406,207 B2 | Claim 1 (Original Text / English)

An artificial intelligence system comprising: at least one memory storing instructions; and at least one processor configured to execute the instructions to perform operations comprising: receiving a query to configure a custom model, the query comprising features, the features comprising a knowledge base and a capability; providing the features to a model builder, the model builder being a language model configured with a specific instruction sequence for the formation of custom models, the specific instruction sequence comprising instructions for the language model to behave as an expert at creating custom models; configuring the custom model using the model builder based on the features by fine-tuning a base model with the knowledge base and generating a set of instructions to configure the capability; generating a custom interface for interacting with the custom model; receiving a prompt via the custom interface; and generating a response using the custom model, wherein the specific instruction sequence comprises: a first step of establishing a behavior for the custom model based on the features; a second step of generating a profile for the custom model; a third step of generating refining prompts for configuring the custom model, the refining prompts seeking parameters for at least one of: data extraction from the knowledge base, a template, an expected output, or an interface configuration; and an instruction to follow the steps in order without skipping any.

Reference Translation by a Patent Attorney (Japanese)

An AI system comprising: a memory for storing instructions; and a processor for executing the instructions to perform the following: • Receive a query for
configuring a custom model.The query includes features (features = knowledge base and capabilities); ・Providing
the features to a “Model Builder.” The Model Builder is a language model composed of a specific sequence of instructions for forming a custom model and is instructed to “act as an expert in creating custom models”;
・The Model Builder fine-tunes the base model using the knowledge base, generates a set of commands that constitute the capabilities, and constructs the custom model; ・Generates a dedicated
interface for interacting with the custom model; ・Receives prompts
via the interface and generates responses using the custom model.
Here, the specific sequence of instructions includes: (1) establishing behavior, (2) generating a profile, (3) generating a refinement prompt (to extract from the knowledge base, specify templates, define expected output, and determine parameters for interface configuration), and (4) instructions to execute each step sequentially without skipping any.

Organization of Limitations Supporting Patentability

Limitations Technical Meaning Reason for Inclusion
Model Builder (Meta-AI) A language model that acts as an expert to create AI The Original Core of the Invention
4-Step Command Sequence Rigorous Structuring of the Creation Process Moving Beyond Abstract Ideas
Fine-tuning + Interface Generation General-Purpose → Specialized + Automatic Generation of Dedicated UI Concrete data processing
Execute in sequence without skipping steps Strictness of Procedures Ensuring Specificity of Processing

In addition to Claim 1, independent claims 9 and 14 have also been filed, providing multi-layered protection from different perspectives such as method and medium.

Two Methods for Creating Agents: “AI-Generated” vs. “Human-Written”

Regarding “how to create custom agents,” OpenAI and Anthropic have obtained patents based on contrasting approaches.

  OpenAI (this article) Anthropic (Previously Discussed)
Patent US 12,406,207 B2 US 2025/0299023 A1
Creator Created by AI (Model Builder) Described by a human using a dedicated language
Core Technology Meta-AI + 4-Step Process + Fine-Tuning Adept Workflow Language + DOM Abstraction
User Experience You convey your requirements, and the AI builds the workflow Describe workflows like code
For example "Asking the AI to build it for you" "Writing the blueprints yourself"

💡 Key Point: Even though both involve “agent construction,” the contrasting approaches—“AI builds interactively” (OpenAI) and “humans describe it using a dedicated language” (Anthropic)—have each been patented separately. This is a prime example demonstrating that there is scope for patent protection based on the specific “method” of construction.

How will this be evaluated in examinations in Japan, the U.S., and Europe?

United States (USPTO)

Rather than being an abstract concept of “building AI with AI,” this patent includes specific processes—such as the four-step Model Builder procedure, fine-tuning, and interface generation—and therefore passed the Alice/Mayo tests. Furthermore, it was registered quickly through Track One (expedited examination).

Japan (JPO)

The application describes specific data processing steps—fine-tuning, command generation, and interface generation—and is structured in a way that makes it easy to satisfy patent eligibility as a software-related invention. The key factors for demonstrating non-obviousness are the “structured creation procedure using meta-AI” and the “effect of reducing computational resources.”

Europe (EPO)

The technical effects of “improved efficiency and reduced computational resources” are clear, making this a configuration that is easily evaluated as a technical feature even under the COMVIK approach.

A detailed comparison of examination practices for AI agent patents in Japan, the U.S., and Europe is provided in “Patent Cases and Examination Practices in Japan, the U.S., and Europe.”

Lessons for Your Own Applications

① Patent the “creation process” as well. Not only the product itself but also the mechanism for creating it (meta-AI, builder) can be patented.

② Strictly structure the procedures. Describing the process concretely—such as “execute the four steps in sequence without skipping any”—enhances patent eligibility.

③ Emphasize technical effects. Effects such as “reduction of computational resources” and “improved efficiency” provide strong support for claims of inventive step.

④ Include safety mechanisms. Incorporating safety checks—such as the Reviewer Model—into the scope of protection provides robust coverage that aligns with actual implementation practices.

⑤ Utilize expedited examination. OpenAI secured rapid registration through Track One. In fast-paced competitive fields, early grant of rights is strategically effective (Japan also has an expedited examination system).

We’ll assess whether your company’s AI creation and customization technologies are patentable.

Patent attorneys with expertise in the IT, software, and AI fields provide comprehensive support—from claim drafting (including creation processes and meta-AI) and free assessments of patentability to utilizing expedited examination and developing filing strategies in Japan, the U.S., and Europe.

Schedule a Free Initial Consultation IT & AI Intellectual Property Services

Frequently Asked Questions (FAQ)

Q. What kind of patent is US 12,406,207 B2?

A. It is a U.S. registered patent held by OpenAI that protects a system for generating users’ own “custom AI models (AI agents).”At its core is a language model called the “Model Builder,” which acts as an “expert in creating custom models” and handles everything from user inquiries (knowledge base and capabilities) to fine-tuning the base model, generating instructions, and creating a dedicated interface.Registered on September 2, 2025, with a total of 19 claims. It supports OpenAI’s “GPTs” feature.

Q. What is a “Model Builder”?

A. It is a language model configured to create custom AI—an “AI that creates AI (meta-AI).” It is structured with a specific sequence of instructions to “act as an expert in creating custom models.” In response to user requests, it follows a procedure involving establishing behavior, generating profiles, and creating refined prompts to build custom agents.

Q. Does this patent refer to “GPTs”?

A. The structure described in the patent specification corresponds to OpenAI’s “GPTs (Custom GPT)” feature. It is a mechanism that allows users to create and deploy unique AI assistants equipped with specialized knowledge and capabilities without writing code, and this patent protects the underlying technology.

Q. How does this differ from Anthropic’s agent-building patent?

A. The underlying philosophy differs. Anthropic’s patent (US 2025/0299023) uses a dedicated language called “Adept Workflow Language” in which “humans describe workflows.”OpenAI’s patent, on the other hand, takes an approach called “Model Builder,” in which “AI creates agents interactively and structurally.” Although both involve “agent construction,” they represent contrasting designs: one where humans write the code, and the other where AI creates it.

Q. Can methods for creating custom AI also be patented?

A. Yes, it can. As demonstrated by this patent, if you describe specific processes—such as the structured procedures, fine-tuning, and interface generation of a meta-AI (Model Builder)—and demonstrate a technical effect such as improved efficiency (reduction in computational resources), it is possible to obtain patent protection in Japan, the U.S., or Europe.

Important Note Regarding This Article: This article provides a general explanation of technology and legal systems based on published patent applications. Although US 12,406,207 B2 is a registered patent, the actual scope of protection is determined by the wording of each claim, the doctrine of equivalents, and historical information.The cited claims, abstract, and specification are based on published patent bulletin data (such as FreePatentsOnline); however, for legally significant purposes (FTO, infringement analysis, invalidity, patent applications, etc.), please be sure to verify the official USPTO transcript and the latest prosecution history, and consult with an expert for a case-by-case review.The Japanese translation is provided for reference purposes only; the official text is the original English version.

Recommended Reading (AI Agent Patent Series):
Anthropic’s “Agent Construction” Patent (Described by Humans Using a Dedicated Language)
OpenAI’s Multi-Agent Patent (Shared Workspace)
OpenAI’s “AI Agents Learning from Videos” Patent (VPT)Patent Case Studies and
Examination Practices in Japan, the U.S., and Europe (Case Studies)

Source