The patents and the accused products
Flinn Artificial Intelligence (AI) Technologies, LLC v. Anthropic, PBC, No. 1:26-cv-01454 (D. Del.), was filed November 12, 2026, as a 62-page complaint for patent infringement, with a jury demand. The complaint describes Flinn as a Texas limited liability company with its principal place of business in Woodville, Texas, and Anthropic as a Delaware public benefit corporation with its principal place of business in San Francisco, California. It alleges that the named inventors assigned the five patents to Flinn, that the assignments were recorded at Reel 76630, Frame 0145, and that Flinn is the sole owner and has not licensed them. The complaint alleges that Anthropic built its business on the architectures recited in the patents, and it seeks compensation for what it calls Anthropic's unauthorized use of Flinn's patented inventions.
The five patents are each titled "Iterative attention-based neural network training and processing." U.S. Patent 12,327,166 was filed November 27, 2024, issued June 10, 2025, and states 397 claims on its face. U.S. Patent 12,223,404 was filed January 26, 2023, issued February 11, 2025, and states 375 claims. U.S. Patent 12,511,580 was filed December 6, 2024, issued December 30, 2025, and states 65 claims. U.S. Patent 12,462,197 was filed December 6, 2024, issued November 4, 2025, and states 128 claims. U.S. Patent 12,462,198 was filed January 16, 2025, issued November 4, 2025, and states 115 claims. The complaint alleges that each patent claims priority to an application filed January 18, 2016, which issued as U.S. Patent 10,510,018, and to an application filed August 3, 2015. The faces of U.S. Patents 12,327,166, 12,511,580, 12,462,197, and 12,462,198 each state that the patent is subject to a terminal disclaimer.
The complaint alleges that Anthropic operates the claude.ai website and the Claude platform. It names the accused products as Claude and its chat, agentic, and task-execution functionality, including Claude Chat, Claude Code, Claude Cowork, Claude Design, Claude Enterprise, Claude API access, and hosted or dedicated Claude deployments. It alleges that those products are powered by models from the Claude series, and it lists Claude 1, Claude 1.2, Claude 2, Claude 2.1, Claude 3, Claude 3.5, Claude 3.7, Claude 4, Claude 4.1, Claude 5, and the Opus, Sonnet, Haiku, Fable, and Mythos versions, together with later-released, successor, replacement, updated, modified, derivative, or renamed products that perform the same or substantially similar accused functionality. The complaint alleges infringement by making, using, offering for sale, selling, and/or importing the accused products.
What the claims cover
The complaint says the claims it discusses are examples, not a complete list, and that the claim charts are illustrative and not exhaustive. Count 1 identifies at least claim 1 of the '166 patent, and Exhibit A is an exemplary chart of that claim. Count 2 identifies at least claims 10 and 21 of the '404 patent, and Exhibit B is an exemplary chart of those claims. Counts 3, 4, and 5 identify at least claim 1 of the '580, '197, and '198 patents, and Exhibits C, D, and E are exemplary charts of those claims. The complaint also quotes claim 3 and claim 373 of the '166 patent, claim 2 of the '197 patent, and claims 274 and 357 of the '404 patent as further examples. The complaint alleges that the claimed arrangements were not well-understood, routine, or conventional as of the August 2015 and January 2016 priority dates, and that the claims do not preempt attention-based neural network processing generally.
U.S. Patent 12,327,166, claim 1, covers a computer-implemented method at a system with access to a trained computer-implemented neural network. Claim 1 covers causing access to information that includes a first plurality of syntactical elements; causing prioritization of a first plurality of attentions, including a first attention associated with a representation of a first subset of those elements and a second attention associated with a representation of a second subset; causing generation, by application of the trained computer-implemented neural network, of a first plurality of probabilities based on that prioritization; causing prioritization of a second plurality of attentions, including a third attention and a fourth attention, based on the first plurality of probabilities; causing generation of a second plurality of probabilities based on the prioritization of the second plurality of attentions; causing generation of a second plurality of syntactical elements based on the second plurality of probabilities; and causing a communication to be sent to a user.
Claim 3 of the '166 patent depends on claim 1. It covers a response perceptible to the user, where the user includes a human user, caused based on a prompt that includes the first plurality of syntactical elements. At least one operation is caused based on the second plurality of syntactical elements to generate a representation of a rationale for those elements, and that representation is used to generate a third plurality of syntactical elements on which the response is based. The operation includes generation, utilizing the trained computer-implemented neural network, of a plurality of sequences of elements, where each sequence includes multiple elements; evaluation of the sequences; selection of at least one of the sequences based on the evaluation; and inclusion of a representation of that sequence with the representation of the rationale. Claim 373, which also depends on claim 1, covers storing the representation of the first subset of the first plurality of syntactical elements and retrieving the stored representation for use in prioritizing one or more other attentions associated with one or more representations of one or more other syntactical elements that are utilized for the generation of the second plurality of syntactical elements.
U.S. Patent 12,223,404, claim 10, covers a system with one or more processors, one or more memories in communication with the processors, and one or more programs stored in the memories to be executed by the processors. The programs include instructions for causing access to a trained computer-implemented neural network of a type other than a recurrent neural network; access to information including a first plurality of syntactical elements; generation, by application of the trained computer-implemented neural network, of a first plurality of probabilities associated with those elements; a first attention directed to a representation of a first subset of the elements, based on the first plurality of probabilities; generation of a second plurality of probabilities based on the first attention and the first plurality of probabilities; a second attention directed to a representation of a second subset of the elements, based on the second plurality of probabilities; generation of a third plurality of probabilities based on the second attention and the second plurality of probabilities; generation of a second plurality of syntactical elements based on the third plurality of probabilities; and a communication sent to a user.
Claim 21 of the '404 patent depends on claim 10. It covers a user prompt capable of including at least one image in addition to and corresponding with the first plurality of syntactical elements, so that the first plurality of probabilities is associated with the image in addition to the elements, and, based on the first plurality of probabilities, one or more attentions directed to a representation of at least a portion of the image in addition to the representation of the first subset. Claim 274, which also depends on claim 10, covers further training of the trained computer-implemented neural network utilizing training syntactical elements and a first plurality of images, so that a first plurality of correspondences is identified between syntactical elements and patterns of pixels; access to a third plurality of syntactical elements that describes a scenario; identification of a second plurality of correspondences between those elements and sets of pixels; and generation of one or more images that represent the scenario by applying a first plurality of attentions based on the second plurality of correspondences. Claim 357, which also depends on claim 10, covers a first plurality of syntactical elements that represents a request for the system to perform an action other than generating content utilizing the trained computer-implemented neural network and other than causing the communication to be sent to the user, performance of that action based on the second plurality of syntactical elements, and a communication that reflects the performing of the action.
U.S. Patent 12,511,580, claim 1, covers a computer-implemented method at a system including one or more processors and one or more memories, with instructions stored in the memories. Claim 1 covers access to a first plurality of syntactical elements and generation, by multiple-attention-based neural network application, of a vector representation of those elements. The claim describes that application as involving a computer-implemented neural network trained utilizing hardware that processes representations of training syntactical elements based on multiple training attentions that correspond to representations of different subsets of the training syntactical elements. Claim 1 then covers a search of stored vector representations, each representing at least one of a plurality of distinct items of content and each generated by multiple-attention-based neural network application; a comparison of the vector representation of the first plurality of syntactical elements with at least a portion of the stored vector representations; identification of one or more of the distinct items of content based on the comparison; generation, by multiple-attention-based neural network application and based on the identified content, of a second plurality of syntactical elements; and a communication, based on the second plurality of syntactical elements, sent to a user.
U.S. Patent 12,462,197 states 128 claims on its face. The complaint identifies at least claim 1 in the fourth count and in an exemplary chart. Claim 2 depends on claim 1. Claim 2 covers a prompt from a user who includes a human user, a communication of a third plurality of syntactical elements to that user before a determination whether a first scenario represents objective reality, and use of the user's response for that determination, such that when the determination is that the first scenario does not represent objective reality, a fourth plurality of syntactical elements is automatically generated based on the response. U.S. Patent 12,462,198 states 115 claims on its face. The complaint identifies at least claim 1 in the fifth count and in an exemplary chart. Claim 1 covers a computer-implemented neural network trained on computer code so that it is configured to generate probabilities in connection with a plurality of executable computer code elements for selection from them; a first attention and a second attention directed to representations of subsets of a first plurality of syntactical elements; generation, by application of the trained computer-implemented neural network, of probabilities that are each associated with a corresponding subset of the executable computer code elements, based on the second attention; selection of one or more of those code elements based on the probabilities; and a communication sent to a user.
The counts, the relief, and venue
The complaint pleads five counts, in order, for infringement of the '166, '404, '580, '197, and '198 patents. Each count is pleaded under 35 U.S.C. 271. Each alleges that the accused products meet all limitations of the claims identified for that count, and that Anthropic directly infringes under 35 U.S.C. 271(a), literally or under the doctrine of equivalents, by making, using, offering for sale, selling, and/or importing the accused products. The counts on the '166, '580, '197, and '198 patents also allege infringement by performing the claimed methods. Each count alleges induced infringement under 35 U.S.C. 271(b) since at least the filing and service of the complaint, with specific intent to induce infringement or, at minimum, with willful blindness. Each count alleges contributory infringement under 35 U.S.C. 271(c), since at least the filing and service of the complaint, by sale and offer of a provisioned and configured Claude deployment, which the complaint describes as the trained Claude model together with the configuration, tuning, serving parameters, and enabled functionality Anthropic supplies. The complaint alleges that the deployment is a material part of the invention, that Anthropic knows it to be especially made or especially adapted for use in infringement, that it has no substantial non-infringing use, and that it is not a staple article or commodity of commerce. Each count alleges that the infringement is willful, and alleges knowledge of the patent and of the element-by-element claim chart since at least the filing and service of the complaint. Each count seeks enhanced damages under 35 U.S.C. 284 and attorneys' fees, expenses, and costs under 35 U.S.C. 285. The complaint alleges that each patent is valid and enforceable and carries a presumption of validity under 35 U.S.C. 282.
On extended thinking, memory, retrieval, and selection of code, the complaint alleges, on information and belief, that accused configurations generate, evaluate, and select among intermediate sequences; store representations generated during attention-based processing and later retrieve them for attention prioritization; generate and compare vector representations of content; and generate probabilities associated with corresponding subsets of executable computer code elements. The complaint cites Anthropic's model cards as describing Claude models as large language models that use a transformer architecture, and alleges, on information and belief, that later Claude models, including Claude 3 and its successors, also use an attention-based transformer architecture. The complaint says that its allegations about the internal operation of the accused models rest on public information and the claim charts, and that it expects them to be confirmed and supplemented in discovery.
The complaint pleads subject matter jurisdiction under 28 U.S.C. 1331 and 1338, and personal jurisdiction because Anthropic is organized and exists under the laws of Delaware. It pleads venue under 28 U.S.C. 1400(b) on the ground that Anthropic resides in this district, because it is a domestic corporation organized under the laws of Delaware. It alleges that Anthropic maintains offices in Seattle, Washington, and New York, New York. It alleges, on information and belief, that Anthropic's employees access, use, test, develop, and deploy the accused products from locations throughout the United States.
The prayer asks for a judgment that Anthropic has infringed and continues to infringe, directly, indirectly, and willfully, one or more claims of each patent; damages adequate to compensate for the infringement, in no event less than a reasonable royalty, together with interest and costs under 35 U.S.C. 284; a judgment of willfulness and enhanced damages up to three times the amount assessed under 35 U.S.C. 284; an accounting of infringing acts and supplemental damages through entry of final judgment, with interest; an ongoing royalty for continued infringement after final judgment; pre-judgment interest from the date of first infringement and post-judgment interest under 28 U.S.C. 1961; costs under 28 U.S.C. 1920 and Rule 54(d) of the Federal Rules of Civil Procedure; a judgment that the case is exceptional under 35 U.S.C. 285, with reasonable attorneys' fees, expenses, and costs; and other relief. The complaint alleges that Flinn does not make, use, offer for sale, sell, or import any article covered by the patents and has not licensed any person who does, that no marking was required under 35 U.S.C. 287, and that Flinn is entitled to recover damages for infringement occurring before the complaint, subject to 35 U.S.C. 286. Flinn demands a jury on all issues so triable.
What to watch next
Watch for Anthropic's answer or any motion under Rule 12 or to transfer venue, and for any identification of asserted claims under a scheduling order. The case can end at any time, including by settlement or voluntary dismissal.