USA Today Sues OpenAI for $250m: What 122 Million Training Tokens Are Worth

The filing, the parties and the demand

The case is docketed as USA Today Co., Inc. v. OpenAI Foundation, No. 1:26-cv-08892, and names OpenAI's corporate entities as defendants. Nineteen titles sit on the plaintiffs' side — the national paper plus regional and local titles published by USA Today Co. and its affiliates. The claims are direct copyright infringement, vicarious copyright infringement, and intentional removal of copyright management information. The last one is a DMCA-style count rather than a licensing dispute.

The publishers ask for at least $250 million in total damages, up to $150,000 per willfully infringed work, and up to $25,000 for each instance where copyright attribution was allegedly stripped. According to the complaint, more than 160,000 of their articles and web-text entries appear in OpenAI's training datasets. The same filing attributes 122+ million tokens of their published output to a 2019 Common Crawl C4 snapshot, including 23 million tokens from usatoday.com.

OpenAI's answer is the same one it has given since 2023: training on publicly available web text is fair use and the resulting model is transformative. The company holds licensing agreements with several large publishers. These plaintiffs are not among them. The Economic Times reported the filing on 8 October.

The arithmetic behind the headline number

$250 million divided by the $150,000 statutory ceiling for a willful infringement comes to roughly 1,667 works. That is about one percent of the 160,000+ entries the plaintiffs say they identified in the training data. If a court were to count every one of those entries at the maximum rate, the theoretical exposure would land near $24 billion. Nobody is asking for that, which tells you the demand is a negotiating position rather than a computed bill.

The attribution claim is easier to price on paper. At $25,000 per instance, the same 160,000 entries would generate about $4 billion in penalties. Again, the complaint does not go there.

122 million tokens at current API prices

Here is a number worth holding next to the $250 million. OpenAI's current API list price for GPT-6.1 Sol, released on 29 September 2026, is $2.00 per million input tokens and $10.00 per million output tokens, with cached input at $0.10 per million. The 122 million tokens the complaint attributes to the plaintiffs would cost about $244 if pushed through that API at the standard input rate. The 23 million from usatoday.com come to $46.

This is not a legal argument and I am not pretending it is one. Training corpora are not billed at inference rates, a Common Crawl snapshot from 2019 is not a licensed feed, and the value of a news archive is not the price of the characters inside it. The comparison only shows the gap between what publishers say the data was worth to OpenAI and what the same volume of text costs when it flows through the product the data helped build. In our own runs on the RTX 5060 Ti in AI Arena, a single corpus evaluation can burn millions of tokens in minutes, which is exactly why publishers argue for a share of the output, not a per-token fee.

Fair use is still not settled

The complaint lands alongside several other copyright actions against OpenAI in the same district, and the USA Today filing adds a national title with a large regional footprint to that group. No ruling has established that training a commercial model on scraped news is or is not fair use. Until an appellate court says otherwise, publishers and AI developers are both operating on assumptions, which is why so many of them are signing licences instead of waiting for a judgment.

Europe works from a different rulebook

The US case turns on fair use. The EU does not have that doctrine in the same form. Text and data mining is permitted under the DSM Directive's Article 4 exception unless a rightsholder has reserved their rights in a machine-readable way, and Article 3 covers research organisations separately. That opt-out is what European publishers have been attaching to their sites and feeds since 2019, and it is the mechanism they would rely on in a European dispute.

There is a second layer now. Since 2 August 2026 the European Commission and the EU AI Office hold full enforcement powers over general-purpose AI providers: mandatory technical evaluations, compelled documentation, and fines of up to 3% of global annual turnover or €15 million, whichever is higher. Providers also publish a summary of their training content under the AI Act framework, and Article 50 transparency duties apply to synthetic content. For scale, the EU's maximum fine on a general-purpose model provider is a rounding error next to a $250 million civil claim in New York.

What a European team should check before shipping

Deployers now share accountability for the models they put into products. If you are building a summariser, an archive search tool or a news assistant on top of a US frontier model, the practical question is no longer only about copyright. You should be able to show, if asked, that your provider maintains technical documentation, mitigates systemic risk and stays current with its AI Act filings. That usually means a contract clause, a documented evaluation routine and a paper trail, not a blog post.

Alternatives exist. Mistral Large 4 entered public preview on 6 October, with open weights promised by the end of October, which gives European teams an option that can be self-hosted and audited. The trade-off is capability, not paperwork, and it is one worth measuring rather than assuming.