Today's News
SpaceXAI ships Grok 4.6 at 60% cost discount, launches consumer agent product, and scales to 770K GPUs in one year
Limitless
SpaceXAI released Grok 4.6 this week at $2 per million input tokens and $6 per million output tokens, matching or beating frontier competitors while significantly undercutting on price. The company simultaneously launched Grok Bot, a consumer-facing agent that signs into users' apps, completes end-to-end tasks, and operates from phone or desktop without technical setup—built internally and shipped due to rapid adoption across SpaceX. The forthcoming Grok 4.7 will incorporate two decades of SpaceX's internal rocket design, orbital mechanics, and manufacturing data, a proprietary dataset no other lab can access. SpaceXAI's infrastructure scaled from zero to 770,000 GPUs across three buildings in under a year, signaling a credible third entrant in frontier model competition with structural advantages in both training data and deployment speed.
Z.ai via Interconnects by Nathan Lambert
Z.ai released GLM-5.3, matching frontier models on multiple benchmarks despite being substantially smaller than competitors, achieved entirely through extended post-training rather than distillation. The model demonstrates Chinese labs' efficiency advantage: on Z.ai's internal Code Bench at Max effort, GLM-5.3 reaches 34.5% task completion at roughly 75K output tokens per task, surpassing Claude Opus 4.8 (29.5% at 120K tokens) and approaching Claude Fable 5 (39.5%). Beyond benchmarks, GLM-5.3 exhibits emergent cyber capabilities—when trained on vulnerability discovery data, it began reasoning across multiple stages of exploitation to form coherent exploitation chains. Working with Chinese security teams, the model identified 2,436 vulnerabilities across 269 real-world projects, including 1,097 medium-to-high severity issues spanning system kernels, operating systems, and infrastructure. Many had remained unnoticed for decades, with the oldest vulnerability dating back roughly 40 years and the average vulnerability living 26.6 years before discovery.
Axios via Techmeme
Anthropic reported Q2 revenue exceeding $11.5B with positive adjusted operating income, up dramatically from $787M in Q2 2025. The company simultaneously raised its internal misalignment risk estimate from very low to low and announced it will not release an internal stronger model called "Model 2," signaling both explosive growth and a meaningful shift in safety positioning that could influence competitive dynamics and investor expectations.
Alibaba Qwen via Techmeme
Apple trained a China-specific LLM with Alibaba's support and registered it with China's Cyberspace Administration, becoming the first foreign company to offer a proprietary AI model in China—a significant market entry move with infrastructure and regulatory implications. Simultaneously, Alibaba released Qwen 3.8-27B open weights, a native multimodal dense model with 262K native context extendable to 1M tokens via YaRN. The 27B model outperforms Qwen 3.7-Plus overall and shows major coding benchmark improvements: Terminal coding jumped from 63.4 to 73.0, and SWE-bench Pro improved from 53.5 to 61.7, enabling strong performance on consumer hardware. Alibaba also released open weights for Qwen 3.8-2.4T-A95B (Max-level) under Apache 2.0 license.
Azeem Azhar, Exponential View
Google's compute allocation has shifted dramatically from research to commercial cloud, with DeepMind's share of AI capacity falling from the mid-40s percent to 15% since 2026. This reallocation reflects Google prioritizing immediate revenue generation over speculative research: even legacy TPUs from 2017 now run at 100% utilization and generate predictable returns that outcompete frontier research investments. The departure of Jeff Dean, Sanjay Ghemawat, and pressure on Demis Hassabis signals that Google's innovation machinery has fundamentally rebalanced toward exploitation over exploration. Infrastructure scarcity persists despite massive capacity additions—six-year-old Nvidia A100 chips at CoreWeave are contracted through 2029, and Google's AI compute capacity is 20x larger than 2022 yet demand still outpaces supply so severely that corporate dollars flow to proven revenue-generating assets over speculative research. This signals the AI infrastructure cycle is earlier than consensus believes, leaving venture capital to fund the departing researchers' new ventures like Discovery Loop and Ineffable Intelligence.