AI News: ChatGPT Ultrafast, Grok 4.6, 3 New Open-Source Models, and more!

Model Performance & Infrastructure - OpenAI launched ultra-fast mode running at 14 to 15 times speed utilizing Cerebras custom chips, significantly reducing development time [2][3][4] - Industry deployment shows large language model bottlenecks shifting from GPUs to traditional CPUs and local computers during tool calling [6][7] - OpenAI released "Computer History" to record and automate computer tasks, offering opt-in and fine-grained privacy controls [29] Regulatory Compliance - Anthropic integrated hidden watermarks into model outputs to comply with EU AI Act Article 52 regarding the transparency of AI-generated content [8][9] - Watermarking utilizes low-stakes choices and random source adjustments during word selection to ensure minimal impact on actual user outputs [12][15][16] Market & Open Source Dynamics - xAI released Grock 4.6 priced at $2 per million input tokens and $6 per million output tokens, powering the new simplified Grockbot agent [18][19] - DeepSeek launched DeepSeek V4 Pro featuring tiered pricing, with off-peak cache misses at $0.66 per million input tokens and $1.98 per million output tokens, and cache hits at $0.02 per million input tokens [25] - Meta returned to open source by releasing Muse Glimmer, a 30 billion parameter model optimized to run on-device using desktop GPUs [26][27]