The AI Trademark Battle: The Biggest Names in AI Don't Own Their Names

Three of the most valuable product names in technology were not cleanly owned by the companies that made them famous. Google rebranded Bard to Gemini in February 2024. A company called Gemini Data had been using the same name in commerce since 2011 and had a federal registration to prove it. xAI shipped Grok in late 2023. Two other companies had registered GROK at the US trademark office years earlier. Anthropic built Claude into a household name. A small software company called Get Claude, LLC had a CLAUDE software application on file before Anthropic's own filings cleared.

None of these AI trademark disputes is rumor or analysis. It is sitting in the public trademark register, where every application, refusal, opposition, and registration is timestamped and searchable. Signa pulled the full filing histories for the four leading AI labs and read them back. The pattern is consistent enough to be a strategy, or at least a shared blind spot: the AI industry ships first and files later, and the register has been keeping score the whole time.

• The biggest AI brands launched on contested names. Gemini, Grok, and Claude each had a senior US trademark owner (someone who got there first) before the AI product existed. Not one of these flagship word marks is cleanly registered to the company that made it famous.
• Gemini is in active litigation. Gemini Data sued Google for trademark infringement in September 2024. The USPTO refused both of Google's Gemini AI applications, and Google's older Gemini smartphone registration is now facing cancellation.
• The trademark office doubles as a launch radar. OpenAI filed DAYBREAK the same day it launched Daybreak. Anthropic filed CLAUDE MYTHOS one day before its Fable 5 launch. Filings track product reveals tightly enough to forecast them.
• The one thing they do lock early is model names. OpenAI has applications on file for GPT model numbers that have no released product behind them. Companies that ship product names late are fastidious about reserving version numbers.

Here is the scoreboard, drawn straight from the register.

Every number below comes from the public trademark register, pulled via Signa and current to mid-June 2026. Statuses change, and several of these files are moving. The methodology and sourcing are at the end.

Gemini: Google launched on a name that was already registered

This is the most documented conflict in the set, and the cleanest illustration of the pattern, so start here.

A Nice class is a category in the international system for sorting the goods and services a trademark covers, numbered 1 through 45. Class 9 is software and electronics. Class 42 is technology services. Almost every AI mark in this article lives in one or both. Keep those two numbers in mind, because the whole Gemini fight turns on Class 9.

Gemini Data Inc., an enterprise data and analytics company, has used the GEMINI mark in commerce since 2011. It holds USPTO registration 6389900, filed on May 8, 2019 in Class 9 and registered in 2021, with no challenges against it. That use predates Google's rebrand of Bard to Gemini, in February 2024, by roughly thirteen years.

Google was not flying blind. It already held an older GEMINI registration (5489150, filed in late 2015 and registered in 2018) for smartphones. But selling phones under a name is not the same as clearing it for a flagship AI assistant in a class someone else already owns. When Google filed two new Gemini applications aimed squarely at the AI product (app 98202646 in Class 42, filed September 28, 2023, and app 98418146 in Class 9, filed February 23, 2024), the USPTO refused both in 2024. The cited ground was likelihood of confusion, the core test US examiners and courts use to decide whether two marks are close enough that consumers would assume they share a source. An earlier mark in the same class for related goods is exactly what triggers it.

Both Google applications are still in examination, not registered. Google requested that examination be suspended pending the outcome of litigation, because on September 11, 2024, Gemini Data sued Google for trademark infringement in the Northern District of California. And Google's own older smartphone registration (5489150) now carries a pending cancellation, a proceeding that asks the trademark office to strike a registration that already issued. Filing a cancellation against the senior party's registration is a standard countermove in a dispute like this.

Read the timeline as a single line and the collision is obvious. Gemini Data starts using the name in 2011. It registers in Class 9 in 2021. Google rebrands to Gemini in February 2024. The USPTO refuses Google's AI applications later in 2024. Gemini Data sues that September. The most valuable AI product launch of Google's year went out under a name another company had owned, in the relevant class, for more than a decade. That is not a paperwork delay. It is a live infringement case over the name on the box.

Grok: shipped on a name registered twice before it

xAI launched Grok in late 2023. By then, GROK had already been registered at the USPTO twice, by two unrelated companies, and neither registration belongs to xAI.

GrokStream, Inc. filed GROK in 2011 and registered it in 2012 (reg 4184452) in Class 35, the class for business and advertising services. Bizly, Inc. holds a second GROK registration (reg 7262816). Both predate xAI's 2023 launch, and both sit in or near the software and services classes where an AI assistant lives. That is the textbook setup for a likelihood-of-confusion problem.

Against that backdrop, none of xAI's US GROK word-mark applications is registered. All are still in examination. The senior GROK marks are precisely the kind of obstacle that keeps a junior application stuck at the examination stage, and on the present record they are doing exactly that.

xAI has been busy at the register in a different way: it has filed a small constellation of personality marks around the Grok brand. There is GORK (app 99179705, in Classes 9, 25, and 42), the deliberate meme misspelling of Grok, filed defensively across software, clothing, and tech services. There is the tagline TO UNDERSTAND. And there is MACROHARD, the jab at Microsoft, filed at the USPTO on August 1, 2025 (app 99314877, Classes 9 and 42). The filing went in about three weeks before Musk announced the project publicly on August 22, 2025. The register knew about Macrohard before the public did.

That head start did not last. On August 11, 2025, ten days after xAI's own filing, Kelly Kulikoff filed MACROHARD for latex condoms (Class 10, app 99975144). His application sailed through, and it registered on April 14, 2026 (reg 8214890). It is the only registered MACROHARD at the USPTO. xAI's application, in the software and tech-services classes that actually matter for an AI venture, is still examining. The classes do not overlap, so this is not a head-to-head conflict so much as a flag planted, but the optics are hard to miss: the first party to actually own a federal MACROHARD registration sells condoms, not chips. This is trademark squatting at its most opportunistic, racing a famous-by-association name to registration in a class the originator left open.

Claude: a senior filer, an opposition, and a toilet

Anthropic's Claude has the messiest supporting cast.

Start with the senior filer. Get Claude, LLC filed a CLAUDE software mark (app 97327851, Class 42) on March 24, 2022, and that application is in allowed status, meaning the USPTO has approved it and it is on track to register once the formalities close. It went in before Anthropic's own CLAUDE filings cleared, and it sits in Class 42, the same technology-services class an AI assistant occupies. A senior, allowed, directly competitive mark is the kind of thing that turns a trademark application into a negotiation.

It gets stranger by class. Kohler Co., the plumbing-fixtures company, filed CLAUDE for toilets, bathtubs, and sinks (Class 11, app 98864403) on November 20, 2024. That one does not collide with Anthropic, because plumbing fixtures and AI software are about as far apart as two classes get, but it is a useful reminder that a common given name like Claude is already spread across the register in places that have nothing to do with AI.

Then there is Anthropic's own house. The ANTHROPIC word mark, the company's actual name, is itself under opposition, the proceeding where a third party files a formal objection to block a pending application from registering. So both halves of the brand, the product name and the company name, are contested on the register at the same time.

The most telling detail is the timing. On June 8, 2026, Anthropic filed CLAUDE MYTHOS (app 99872437 in Class 9 and app 99872440 in Class 42). One day later, on June 9, 2026, it launched Claude Fable 5 publicly. (Claude Fable 5 is the public product; the related Mythos line is a separate, restricted release.) A trademark application filed one day before a product reveal is not a coincidence. It is the legal team racing the press release, and the register caught the moment on the way past.

OpenAI: the disciplined one (mostly)

OpenAI is the counterexample that proves the pattern. Its filing behavior is the most disciplined of the four, and where it slips, the slip is documented just as clearly.

The discipline shows up in timing. OpenAI filed DAYBREAK (app 99815477) on May 11, 2026, the same day it launched its Daybreak security product. Same day, not weeks after. When a company files on launch day, it has cleared the name before shipping rather than after. That is the opposite of the Gemini and Grok approach, and it is why OpenAI is not the one getting sued over its product names.

It is not flawless. When OpenAI refiled CHATGPT at the USPTO on September 29, 2025 (app 99419666), it widened the net well beyond the original software classes, adding Classes 35, 38, 41, and 45, which cover advertising, telecommunications, entertainment and education, and personal and social services. That is a brand reaching to fence off territory it does not yet operate in, and broad filings invite challenges. The US refile is in examination.

The register also keeps the receipts on what did not work out. OpenAI's SORA application (app 98405663) is under opposition. And GPT STORE (app 98268934) was abandoned, a small headstone for a product direction that did not stick. Abandonment is its own kind of signal: it marks the moment a company stopped paying to defend a name.

Even the disciplined filer gets squatted. State Bull Company Limited holds a CHATGPT registration for tobacco (Class 34) at the EUIPO, the European Union's trademark office, filed before OpenAI locked that class. Tobacco has nothing to do with chatbots, which is the point. Class squatting works by claiming the famous name in a corner its originator forgot to cover.

On the model names, OpenAI is the most forward-looking of all. It has applications on file for GPT version numbers with no released product behind them, names reserved years ahead of any launch. The company that fences off future version numbers is the same one that files product names on launch day. When it matters to OpenAI, it gets there first.

The exception: cooperation, not conflict

Almost every file above is a fence: a company staking out a name against the world. There is one conspicuous exception, and it points the other way.

The MODEL CONTEXT PROTOCOL and MCP marks are not held by a single owner trying to wall off a name. They are split, deliberately, between Anthropic, PBC and the Linux Foundation's LF Projects, LLC. The MCP initialism is registered as two separate marks, both issued on October 14, 2025: Class 9 to Anthropic (reg 7983411) and Class 42 to LF Projects (reg 7983412). Anthropic holds the software-class mark; the Linux Foundation holds the services-class mark for the protocol as an open standard.

That split is not an accident or a clerical artifact. It is a documentary signal of open governance, a structure that says the protocol is meant to be a shared standard rather than one company's property. In a register full of companies racing each other and the squatters to the same names, the one place the AI industry chose to cooperate left a paper trail too.

What the AI trademark disputes reveal

Three findings hold up across all four companies, and each is traceable to the filings.

The trademark office is a near-real-time launch radar. DAYBREAK was filed the day of launch. CLAUDE MYTHOS was filed the day before launch. MACROHARD was filed three weeks before its public announcement. New applications surface on the register in a tight window around product reveals, often slightly ahead of them, which makes trademark monitoring of a competitor's new filings one of the earliest public signals you can get about what they are about to ship.

Ship-first-file-later created real legal liability. This is the load-bearing finding. Gemini and Grok did not just have paperwork pending. They launched on names with senior owners already registered in the relevant classes, and the consequences are on the record: refusals, an opposition, cancellations, and in Gemini's case an active infringement lawsuit. Launching before you have cleared the name is a choice, and for the two highest-profile launches here it bought a documented legal fight.

The one thing they do lock early is model names. The same companies that ship product names late reserve version numbers early, as OpenAI's GPT model-name applications show. Naming discipline at these labs is inverted from what you might expect. The marketing name that the public actually sees is the one most likely to be contested, while the internal version number is locked years ahead.

The cleanest lesson sits underneath all three. The difference between OpenAI's launch-day filings and the Gemini and Grok lawsuits is the difference between clearing a name before you launch and clearing it afterward. The register does not care how good the product is. It cares who filed first.

About the data

Every trademark fact in this article comes from Signa's trademark database, which mirrors the public register: application and registration numbers, filing dates, Nice classes, owner names, and current status (examining, allowed, registered, abandoned, opposed, or facing cancellation). The records were pulled in mid-June 2026. Trademark status is not static, and several files referenced here are actively moving, so a status accurate today may shift as examinations close, oppositions resolve, and the Gemini litigation proceeds. Where a status matters, it reflects the register as of the pull date.

Product launch dates and litigation details were cross-referenced against public reporting: company announcements and blogs, major business and technology press, and legal-industry coverage of the Gemini Data v. Google proceeding. Trademark facts and product facts are kept separate by design. A filing date is what the register says. A launch date is what the company and the press say. The story is in how the two line up.

This is reporting on the public record, not legal advice. Trademark status, likelihood-of-confusion outcomes, and the merits of any opposition, cancellation, or lawsuit are decided by the relevant office or court, not by this analysis. Consult a trademark attorney for guidance specific to your situation.