A new flagship model. A database adding six new hallucinations a day.

OpenAI launched GPT-5.4 on Thursday. It's a competent release, more reliable than GPT-5.3 on long-form coding, faster on multi-step planning, slightly better on chain-of-thought transparency. The product post is well-written. It does not mention the thing happening on the other side of the calendar this week — that Damien Charlotin's database of AI-hallucinated legal citations submitted to courts now stands at 1,227 cases, growing by five to six new ones a day, which means somewhere around forty more were added the week GPT-5.4 came out.

These are not edge users. The Charlotin database is dominated by working lawyers — solo practitioners, mid-market firms, and a steadily increasing share of biglaw associates. They are using ChatGPT and Claude the same way the GPT-5.4 launch post assumes you will use ChatGPT: as a research assistant. The provider is correct that the model is competent at the task. The court filings are correct that the model is not what the lawyer thought it was.

The gap is structural. The product post is a posture document — what the model does well, what features it ships with, what is new since the last release. The court filings are a measurement instrument — what the model produced, what it cited, what was true. There is no provider post in the world that closes that gap, because the provider does not measure what the courts measure. The reason for an independent wire is that the wire is the only thing that does.

On the same calendar, OpenAI published a thoughtful post titled "Reasoning models struggle to control their chains of thought, and that's good," which is the most honest acknowledgement of model limitations in any provider blog this quarter — and a case study from Balyasny Asset Management on building an AI research engine. The week is not lopsided; OpenAI is doing serious work. The point is that the courts are doing serious work too, and only one of those is paid to talk about it.