Three provider stories set the day. Anthropic names its first embedded evaluator. Newly unsealed court filings put a Microsoft scientist on the record about training data. xAI ships a speech to text upgrade while the Grok 4.8 rumor stays quiet.
In the news
Anthropic and Accenture build an embedded evaluation team
Anthropic and Accenture announced a partnership on September 18.1 Faculty, Accenture's specialist AI business, will embed evaluators inside Anthropic to red team models and run alignment assessments.2 Each company expects to invest at least $1 billion over five years.1 The work is non exclusive and will include dialogue with METR and other nonprofits.1 TechCrunch covered the announcement the same day.3
NYT v OpenAI Microsoft unsealed filings quote Microsoft scientist
TechCrunch reported on September 17 that newly unredacted filings in the NYT v OpenAI Microsoft case quote Microsoft director Brent Hecht.1 Hecht called the scraping "the largest theft of labor in human history" in a January 2023 memo.1 The Times said much of the new information comes from its own brief, not the underlying exhibits, which remain sealed.1 The NYT published coverage of the filings on September 17.2
xAI ships Grok Voice Transcribe 2.0
xAI posted Grok Voice Transcribe 2.0 on September 18.1 The release is speech to text, not a Grok 4.x large language model, and the latest LLM post on x.ai remains Introducing Grok 4.6 from August 12. The model offers batch at $0.10 per hour and streaming at $0.20 per hour with word level timestamps and speaker diarization.1 xAI claims the model is twice as accurate as 1.0 and cites Artificial Analysis ranking it first for accuracy among 32 streaming models.1 xAI reports short phrase WER improved from 20.6 percent to 6.8 percent.1
