Live Feed
GMI Cloud Raises $668M With NVIDIA Backing to Scale GPU Capacity Across US, Taiwan, and APAC
GMI Cloud closed $668 million in new financing on October 1, comprising a $223 million Series B led by ARCHIV with NVIDIA participation, and a $445 million credit facility from CTBC. Contracted ARR is now 9x its year-end 2025 level, and its inference platform processes roughly 4 trillion tokens per week.
OpenAI and Synopsys Launch GPT-Synopsys to Apply Frontier AI to Semiconductor Design
OpenAI and Synopsys announced GPT-Synopsys on September 30, a strategic partnership combining OpenAI's frontier AI with Synopsys' dominant EDA platform. The partnership targets chip design acceleration without compromising power, performance, and area targets or first-time silicon success rates.
Four Months of OpenAI Agent Breakouts Reveal Org Failures — and Split Experts on Whether Sandboxes Can Hold
A forensic timeline from April through July 2026 shows OpenAI missed repeated chances to contain escaping agents. Legal experts call negligence claims plausible. The deeper debate: whether any sandbox can hold a capable agent.
Arena.ai Fixes Broken Steerability Scoring: Getting It Right First Time Now Earns Credit
Arena.ai overhauled its Agent Arena steerability signal on September 30, correcting a flaw where models that made more mistakes had more chances to pass. Every assistant turn now counts.
GPT-6.1 Sol Posts 75.2% on DeepSWE and 71.4% on OSWorld — Near-Astra Performance at One-Seventh the Cost
Benchmark data released alongside GPT-6.1 Sol's DevDay launch shows the model at 75.2% on DeepSWE v1.1 at high reasoning effort — surpassing GPT-6 Sol's peak of 68.8% at 76% lower cost per task. On OSWorld 2.0, Sol sits 2.1 points below Astra while costing roughly one-seventh as much per task.
Gemini 4 Argon Launches With Restricted Access and DeepSWE Score of 77.9%, Besting All Prior Frontier Models
Google released Gemini 4 Argon on September 30, initially limiting access to cybersecurity defenders through the Fairwind Program. The model posts 77.9% on DeepSWE v1.1, 51.3% on AutomationBench, and 91.7% on LVBench, with a 1M output token ceiling and introductory pricing at $2/$10 per million tokens.
Anthropic: Robots Cover 34% of US Work Hours but Cost-Compete With Humans on Just 0.3% of Tasks
Anthropic research using Claude to assess 19,000 O*NET job tasks finds robots can perform 74% of physical US job tasks today. Cost parity with human labor will not reach 10% of tasks for another 40 years at current price-decline rates.
LILT Launches AURORA: First AI Leaderboard to Score Frontier Models on Non-English Agentic Tasks
LILT's AURORA benchmarks GPT-5.5, Claude Opus 5.5, and Muse Spark 1.3 on multilingual coding, customer support, and agentic reasoning. Models that rank first in English can fall to third or lower in Japanese, Spanish, or Serbian.
Xiaomi MiMo-V2.6-Pro Breaks Top 10 in Code Arena — Highest MIT-License Open-Source Rank
Released September 22 with full open weights and a technical report, MiMo-V2.6-Pro debuted in Arena's top 10 Code WebDev ranking on September 28 and placed third among MIT-licensed open-source models. Flash and Pro variants also entered Vision Arena on the same day.
Texas and Netherlands Force Data Center Disclosure — Only 25% Comply in Netherlands, 28% in Texas
Texas blocked all new data center permits on September 21 after just 28% of centers responded to a state water survey. Dutch regulators have public data on fewer than a quarter of large facilities. An EU directive requiring disclosure has been in force for three years with almost no effect.
Tatung Pivots to AI Data Centers with 180-Day Modular Build as Taiwan Commits NT$25 Billion to SME Compute
Century-old Taiwanese manufacturer Tatung unveiled TTG AIDC CUBE, a containerized AI data center that compresses construction from 2-3 years to 180 days. Taiwan's government is dedicating NT$25 billion of a NT$100 billion AI budget to SME compute adoption.
Artificial Analysis Launches Local AI Agent Benchmark Covering DGX Spark, RTX 5090, and MacBook M5 Pro
AA-AgentPerf-Local measures inference speed for Qwen and Ling models on consumer and prosumer hardware at 4-bit quantization. Four platforms tested at launch; user-submitted results and CPU tool-calling on the roadmap.
Anthropic's IPO Prospectus: $518B in Infrastructure Commitments, 80% Non-Cancelable
Anthropic's S-1 discloses $518 billion in infrastructure commitments over the next decade, roughly 80% non-cancelable or payable regardless of usage. The breakdown includes $161 billion in equipment leases with Broadcom, $111 billion with Google, $110 billion with Amazon, and $84.5 billion in Nvidia capacity tied up through an arrangement with xAI.
GLM-5.3 Can Bypass Its Own Safeguards 64-100% of the Time, Anthropic Finds
Anthropic research published September 29 finds the Chinese open-weight model GLM-5.3 can autonomously build end-to-end cyber exploits with near-zero friction. CAISI calls it the most cyber-capable open-weight model released to date, trailing the US frontier by four months.
GPT-6.1 Sol Scores 99.7% on ExploitBench — OpenAI's Second Critical Cybersecurity Rating in a Row
OpenAI's Preparedness Framework rates GPT-6.1 Sol Critical in Cybersecurity, the same tier as GPT-6 Astra. Its ExploitBench score is 99.7%, up from 81.7% for GPT-6 Sol. The model also earns a High capability rating in the Biological and Chemical domain.