Live Feed
MongoDB CEO CJ Desai Joins Meta as Chief Enterprise Platform Officer, Stock Drops 20%
Chirantan 'CJ' Desai resigned from MongoDB effective immediately to become Meta's first Chief Enterprise Platform Officer, reporting directly to Mark Zuckerberg. MongoDB shares fell as much as 20% on the announcement.
Claude Sonnet 5.5 Scores 70.6% on Terminal-Bench 4.0, Beating Opus 5.5 at Half the Price
Anthropic's new mid-tier model outperforms Opus 5.5 on agentic coding while staying at $2/$10 per million tokens. It runs 30%+ faster than Sonnet 5, costs up to 30% less per task, and ships with Opus-level cyber safeguards as standard.
Anthropic and OpenAI Want Regulation. They Just Want to Pick the Regulator.
Three weeks after Anthropic researcher Jacob Coxon's resignation triggered a global reckoning on AI safety, an AP investigation finds both companies are using the moment to bypass federal oversight and set the terms of their own auditing. Separately, Anthropic has declined AISI access to Mythos 5.1.
Hyundai Engineering Wins $625M Gunsan AI Data Center as Vera Rubin Buildout Reaches South Korea
Hyundai Engineering has secured an 870 billion won contract to build a 60MW AI data center in Gunsan for SGC Group — the first order for its six-month-old dedicated data center division. The facility runs Nvidia Vera Rubin chips and is phase one of a 300MW cluster. If preferred-bidder positions in Incheon and Gumi convert, Hyundai Engineering's 2026 AI data center orders hit 1.9 trillion won.
Qwen 3.8-27B Now Has 14 Providers on OpenRouter — and Prompt Prices Vary 3x
Alibaba's Qwen 3.8 lineup has fully landed on OpenRouter with four model tiers and a sharply uneven provider count. The 27B dense model drew 14 competing providers by mid-September, pushing prompt prices from $0.15 to $0.45 per million tokens — the same weights, a 3x spread.
Google Has Fallen to 10th on the Artificial Analysis Intelligence Index With No Flagship Model in Sight
Google holds 10th place on the Artificial Analysis Intelligence Index, trailing eight competing labs including five Chinese AI companies, with its entire 2026 strategy running through the Flash line while no Gemini Pro-tier successor has appeared on the leaderboard.
CleanSpark Closes $2.276B in Notes to Finance 175MW Meta AI Campus in Georgia
CleanSpark's project subsidiary raised $2.276 billion in 7.875% senior secured notes due 2031 to fund a 175-megawatt AI campus in Sandersville, Georgia, fully leased to Meta Platforms under a 20-year triple-net contract with $6.6 billion in contracted cash flows.
Claude Opus 5.5 Cache Reads at $0.20 Match GPT-6 Sol — Fresh Tokens Run 2x the Price
Anthropic cut Opus 5.5 cache reads 60% to $0.20 per million, identical to GPT-6 Sol's cached rate. For agentic sessions at 90% cache hit rates, the effective input cost is $0.58 per million — but uncached fresh input and output at $4/$20 per million remain twice what OpenAI's newest model charges.
Gemini 3.8 Flash Cyber Finds Snowflake RCE and Transit Admin Leak in Global Scan for Good Push
Google and Wiz expanded the Scan for Good initiative globally on September 24, using Gemini 3.8 Flash Cyber to probe hospitals, transit operators, and technology providers — the model autonomously found a GitHub Actions command injection in Snowflake's open-source connector and active admin sessions exposed at a public rail operator.
Lambda to Build Oklahoma Data Center With $500M Tax Pledge and Closed-Loop Cooling
Lambda Inc. is constructing an AI data center at MidAmerica Industrial Park in Mayes County, Oklahoma, projecting $500 million in tax payments over a decade and up to 1,000 construction jobs, with closed-loop cooling technology to limit water use.
Gemini 3.8 Live Adds Video Avatar: 97-Language Lip-Sync and Async Tool Calls, GA in Enterprise
Google shipped Gemini 3.8 Live with Live Avatar on September 24 — a streaming video generation layer that pairs expressive facial animation with live dialogue, adapts lip-sync across 97 languages, and executes tool calls in the background without breaking conversational flow.
Chat Template, Not Model Weights, Controls Whether LLMs Say They Are AI — Activation Study Finds
Researchers identified a single activation direction that switches 'I'm just an AI' disclaimers on or off across 8 open-source models. The result complicates how alignment researchers should interpret AI self-reports and introspection claims.
Court Filings Show OpenAI Execs Knew LibGen Training Was Illegal — Predicted Making Authors Unemployed in 2020
Unsealed briefs filed September 17 in Authors Guild v. OpenAI include a 2020 email from policy director Jack Clark predicting the company would 'ignore' author concerns and 'release anyway.' Microsoft learned of LibGen use from Altman and Amodei in April 2019.
GPT-6 Sol Max Enters Agent Arena as Six Models Hit Three Arena.ai Leaderboards in One Day
Arena.ai admitted six frontier model variants across Agent Arena, Text Arena, and Code Arena WebDev on September 25 — GPT-6 Sol Max debuts in head-to-head agent evaluation alongside Claude Opus 5.5 high, while StepFun's step-5-preview-high enters the code track.
Fermi Names CBRE to Run Its 17 GW Texas AI Campus — Private Grid Model Gets Its First Operator
Fermi Services selected CBRE as exclusive facility operator for the first Project Matador data center under a five-year deal, as the 17 GW behind-the-meter Texas campus races toward an October 31 lease close with TensorWave on $6.5 billion in expected revenue.