Flinn AI Technologies Sues Anthropic and OpenAI Over Attention-Based Patents Filed Before Transformer Paper

Flinn Artificial Intelligence (AI) Technologies LLC filed two patent-infringement complaints in the U.S. District Court for the District of Delaware on November 12, naming Anthropic PBC and OpenAI OpCo LLC as defendants. Each complaint alleges infringement of five patents related to neural-network technology, with the accused products identified as Anthropic's Claude and OpenAI's ChatGPT. The suits seek unspecified monetary damages. No rulings, damages figures, or trial dates have been reported [1].

The case numbers are 1:26-cv-01454 (Anthropic) and 1:26-cv-01455 (OpenAI). Sequential numbering suggests the filings landed together. Filing parallel suits is standard when defendants are unrelated, keeping each company's products, evidence, and defenses separate [1].

Patents Predate the Transformer Architecture

The complaints name Steven Flinn and Naomi Moneypenny as inventors. According to Reuters and CBS News, the underlying patent applications date to 2015 and 2016; one report gives the specific filing dates as August 3, 2015, and January 18, 2016 — before Google researchers published the transformer paper in 2017. The transformer design, built around attention mechanisms, became the foundation of modern large language models [1].

Flinn alleges the patents cover "attention-based neural-network architectures." The characterization comes from the complaint text; the patent claims have not been publicly released. If the claims are valid and read on attention broadly enough to cover current models, they would touch nearly every frontier lab. The defendants will likely argue that attention mechanisms existed in research literature before 2015 and that the word "attention" in a patent claim does not automatically cover every system using it [1].

US utility patents generally last 20 years from the earliest filing date, so 2015–2016 applications would remain in force into the mid-2030s, covering the commercial life of today's models [1].

Flinn Ties Patents to Core Revenue Features

The complaints allege that Anthropic's and OpenAI's businesses rely on "reasoning, retrieval, memory, coding, agentic, and multimodal capabilities" covered by the patents. That list maps directly to revenue-driving features: coding and agentic functions drive enterprise subscriptions and API usage; retrieval and memory underpin user retention; multimodal covers image, audio, and document handling. By naming the features customers pay for, Flinn ties alleged infringement to the commercial core of both firms — a framing that matters for damages calculations [1].

Anthropic's IPO Filing Preceded the Suits by Two Days

Anthropic filed an amended S-1 registration statement on November 10, 2026, proposing an initial public offering of 801,351,354 Class A shares at an estimated $70–$78 per share, alongside a $26 billion concurrent private placement of Class N shares from seven strategic partners including Advanced Micro Devices, Inc., Broadcom Inc., Google LLC, Micron Technology, Inc., Microsoft Corporation, NVIDIA Corporation, and TSMC Partners, Ltd. [2][4]. The prospectus, dated November 10, discloses $17 billion in quarterly revenue (three months ended September 30, 2026) and approximately $900.9 billion in commitments, substantially all for compute and infrastructure [2][4].

Prospectus-style disclosures must describe material litigation. If the offering proceeds, these cases would likely appear as risk factors. The financial exposure cannot be sized today because Flinn has not stated a damages number. Patent damages typically rest on lost profits or a reasonable royalty, but no anchor for an estimate has been provided [1].

What Is Confirmed and What Is Not

Confirmed: two complaints exist; filed November 12, 2026; five patents alleged per complaint; inventors Flinn and Moneypenny; case numbers as above; accused products Claude and ChatGPT; unspecified monetary damages sought.

Unconfirmed: no court has found infringement; no damages figure, settlement, injunction request, trial date, or ruling; no company response has been reported; the patent numbers and claim language have not been published; whether the two sets of five patents overlap is unknown [1].

Typical Delaware Patent Timeline

Under federal rules, defendants typically have 21 days after service to answer or move to dismiss (frequently extended). Defendants may petition for inter partes review at the Patent Trial and Appeal Board within one year of service — a common tactic in software cases that can cancel claims outright. Claim construction follows months later, then discovery (source code, financial and technical records), summary judgment, and trial or settlement — commonly well over a year from filing [1].

Courts sometimes pause cases while an inter partes review runs. Enhanced damages for willful infringement (up to three times the base award) are available but require showing the defendants knew of the patents and acted anyway — a high bar for a case this new [1].

Validity Will Likely Dominate

A patent is enforceable only if its claims were new and non-obvious on the filing date. Anthropic and OpenAI can mine pre-2015 research to argue attention-style methods already existed or that claimed combinations were obvious. Flinn must read its claims broadly enough to catch a 2026 chatbot yet narrowly enough to survive prior art — what patent lawyers call "the squeeze," where many cases against fast-moving technology break down [1].

Eligibility is another hurdle. Courts have rejected patents that amount to abstract ideas implemented on generic computers. Software and machine-learning claims draw that challenge regularly, and defendants can raise it early via motion, potentially ending a case before expensive discovery [1].

Industry Implications

If the Flinn patents read on attention-based architectures, exposure extends beyond two defendants. Every lab shipping a transformer-style model — cloud providers, chip-adjacent model makers, open-weight developers — faces similar exposure in theory. Open-weight models raise a distinct puzzle: a developer publishing weights does not sell a service, so damages are harder to compute, but downstream companies hosting those models do earn revenue [1].

Hardware makers have their own angle. Method claims could reach software running on GPUs, TPUs, or NPUs, making the chip vendor's customers, not the vendor, the likelier defendants [1].

Plaintiffs often choose market leaders first. A win against the best-funded defendants sets a reference point for everyone else; a loss may discourage follow-on suits [1].

Historical Context

New platforms attract patent litigation once profitable. The smartphone wars of the early 2010s are the standard example. Generative AI is following the same arc: early disputes focused on copyright and training data; patent claims about underlying architecture are a newer front. One lesson from earlier waves: patent holders with older filings can wait, watching a technology mature, then sue once defendants have deep revenue and few redesign options. A 2015 filing landing in a 2026 complaint fits that profile, though the reports do not say what Flinn did in the intervening years [1].

What to Watch

• First docket entries for service dates, extension requests, and any motion to transfer or dismiss — these will reveal defendants' strategy sooner than press statements.
• Whether Flinn pleads willful infringement.
• Inter partes review petitions, which could cancel claims and pause the cases.
• Any motion on patent eligibility (abstract idea), which could end a case early.
• Disclosure of the patent numbers and claim language, enabling comparison with public descriptions of Claude and ChatGPT.

For now, the filings are allegations, not findings. A complaint is one side's account, and the defendants have not yet answered in any reported filing [1].