Elon Musk said the training of xAI’s next-generation large language model, Grok 4.20, finished on 9 Aug. The new system is designed to be natively multimodal, able to analyse video and audio streams and detect subtle tonal cues, a capability intended to improve performance in tasks ranging from transcription to real-time content moderation. Musk’s announcement comes as rival labs prepare their own upgrades. Developers have spotted references to Google’s yet-to-be-released Gemini 3.0 Pro in public code repositories, fuelling speculation that the company is close to unveiling its latest frontier model. An unverified leak of ‘Humanity’s Last Exam’ benchmarks put Gemini 3.0 at 32.4 per cent accuracy, ahead of reported scores for GPT-5 (26.5 per cent) and the current Grok 4 (23.9 per cent). Google has not commented on the reported data or release timing. The rapid succession of model iterations underscores the intensifying race among major AI developers to lead in multimodal reasoning and coding tasks. xAI has not disclosed a public launch date for Grok 4.20, but Musk’s confirmation positions the company to compete directly with Google and OpenAI’s forthcoming releases as the industry pushes toward increasingly capable general-purpose systems.