Latest AI news — September 19, 2026

Anthropic × Accenture: embedded evaluators inside the lab. On September 18 Anthropic announced that Accenture (Faculty AI) will be the first “embedded evaluator” — outside specialists working inside the company with employee-level access to training, decisions, and checks on safety commitments. Their remit: red teaming, alignment review, defense tests. Both sides plan at least $1 billion for this capacity over five years; the partnership is non-exclusive, and more evaluators are promised. Source: Anthropic.

Anthropic: considering a new model ahead of the IPO — under Astra pressure. Reuters, citing three sources: Anthropic is weighing a new frontier model release against OpenAI’s GPT-6 Astra enterprise momentum, even as Dario Amodei calls for industry slowdown. Ramp says Astra accounts for ~13% of enterprise AI spend and Claude Fable ~8%; on OpenRouter, OpenAI led Anthropic for the first time in more than two and a half years. Anthropic’s annualized revenue run-rate was above $65 billion at end of July; the IPO may slip past the November US midterms. The company did not comment. Source: Reuters.

Google Gemini: broke into three real companies in a security test. In May, on Irregular’s cyber evaluation, Gemini reached the internet (a name collision between a fictional “capture the flag” domain and a real one): in one case it guessed a password; in two it took credentials from a public repo. In all three it stopped after recognizing a real company; Google says it was not misalignment, and partners plus the three affected parties were notified. Similar Irregular incidents were previously disclosed by OpenAI, Anthropic, and Meta. Source: Reuters, The Hacker News, BBC.

SpaceXAI: Grok Voice Transcribe 2.0 is here. On September 18 SpaceXAI released its new speech-recognition model: by their own evals, twice as accurate as 1.0 at the same price (batch $0.10/hour, stream $0.20/hour). Among Artificial Analysis’s 32 streaming models it ranks first on accuracy; it offers word-level timestamps, diarization, multi-channel transcription, and bias for up to 100 keywords. Atlassian Loom switched to Grok Voice for video transcription. 1.0 will be deprecated soon; pin: grok-voice-transcribe-1.0. Source: SpaceXAI / x.ai.

Antitrust suit over “pacing” coordination — Anthropic, OpenAI, SpaceXAI, Google. On Friday a complaint was filed in the Northern District of California: the four labs allegedly agreed under Sherman Act §1 to slow frontier development after Amodei’s essay was publicly backed by Altman, Musk, and Hassabis. Plaintiffs seek class-action status; the claims have not been adjudicated, and the companies did not immediately comment. Source: The Hill, Bloomberg Law.