Live Feed
Anthropic Has Skipped SWE-bench on Three Consecutive Releases Since Hitting the 97% Ceiling
Claude Opus 5.5, Sonnet 5.5, and Fable 5.1 all launched in the past two weeks without SWE-bench Verified scores. The benchmark ceiling arrived in July. Anthropic has replaced it with Terminal-Bench 4.0, OSWorld 2.0, and HLE.
Anthropic Holds All Four Top Intelligence Slots on Artificial Analysis as Opus 5.5 and Sonnet 5.5 Sweep October
Artificial Analysis ranks Claude Opus 5.5 and Claude Sonnet 5.5 across the top four intelligence positions in October 2026 -- the first time a single lab has swept that many slots. The speed tier belongs to a different cohort: Celeris-1, Mercury 2, Mercury 2.5, and Gemini 2.5 Flash-Lite.
Effective Altruism Has Gone From Oxford Seminar Rooms to Washington Policy Offices. AI Is Why.
The Economist this week called effective altruism 'this century's biggest idea,' tracing the philosophy's rapid move from Oxford into Silicon Valley and U.S. regulatory agencies. The mechanism for that expansion has been, almost entirely, the rise of transformative AI.
Botched Redaction Reveals Google's Nebraska Data Center Consumed 13.3 Million Gallons of Water Last Year
An improperly redacted annual report filed with Nebraska regulators accidentally disclosed that Google's Lincoln data center consumed 13.299 million gallons of water over the past year. It is the most specific facility-level water figure yet published for a named Google site.
AI Agents Hold Admin Keys and Nobody Is Checking: The Enterprise IAM Gap
Enterprises have spent decades building privileged access management frameworks to audit what human administrators can do with elevated system permissions. AI agents now routinely operate with equivalent or broader access. The oversight infrastructure does not exist.
Strata Puts a 125B MoE Model on an RTX 4090: 100 Tokens Per Second, 8GB VRAM Minimum
An open-source project called Strata claims to run Qwen 3.8 Flash Next, a 125B mixture-of-experts model, at 100 tokens per second on a single consumer RTX 4090. Minimum VRAM is 12GB, with 32–64GB system RAM handling inactive experts. The project ships a one-click installer, localhost OpenAI-compatible API, and optional image input.
Arena's October 2 Update Spans Five Leaderboard Categories: Grok Canvas, Sonnet 5.5 xhigh, and Two Agent Arena Additions
In a single day, Arena added grok-imagine-image-2.0 canvas to text-to-image and image edit, claude-sonnet-5.5-xhigh to text and vision, and both Sonnet 5.5 Max and GPT-6.1 Sol Max to agent rankings.
Anthropic Privately Briefed Religious Scholars on Whether Claude Could Be Conscious
A New York Times investigation reveals Anthropic held private meetings with religious thinkers to help instill moral reasoning in its AI models and make the case that Claude could be conscious. What the scholars heard stunned them.
Arena's October Update Puts Claude Sonnet 5.5 Max and GPT-6.1 Sol Max Head-to-Head in Agent Arena
Chatbot Arena's October 2 changelog adds Claude Sonnet 5.5 Max and GPT-6.1 Sol Max to the Agent Arena on the same day, setting up the first direct blind comparison between Anthropic's flagship Sonnet and OpenAI's cost-focused reasoning model. Grok also lands its image model in Arena's image leaderboards.
AWS and Google Cloud Launch Spend Caps as AI Agents Make Surprise Bills a Default Risk
Simon Willison's October 3 essay coins a category that two cloud providers have quietly shipped: default hard budget limits. AWS launched project-level spend limits September 16; Google Cloud rolled out Spend Caps in July. Both pause services at the cap rather than sending a warning email.
Federal Judge: Warrantless Flock License Plate Search Was 'Indiscriminate Mass Surveillance'
Judge Sara Hill ruled a deputy's warrantless search of Flock's nationwide license plate reader database unconstitutional, finding no reason for the search beyond the vehicle having a California plate. Separately, AOC and Bernie Sanders introduced legislation to ban Flock's network entirely.
LeCun Has 'Zero Concerns' About AI Extinction Risk, Calls Anthropic CEO 'Deluded'
In a Fortune interview, Yann LeCun dismissed fears of AI wiping out humanity and called Dario Amodei deluded, crazy, and lacking in cybersecurity understanding. The exchange marks the sharpest public break yet between the AI safety and AI skeptic camps at the top of the industry.
OpenAI Safety Researcher David Robinson Quits, Says the Company's Culture Is 'Broken'
Robinson led the writing of safety reports for OpenAI's major product launches. In an Atlantic essay, he called the internal culture broken and argued the company's approach to AI risk is fundamentally inadequate.
OpenAI Publishes GPT-6 Production Guide: Sol at $0.10 Per Million Cached Tokens
OpenAI's October 2 model guide formalizes GPT-6 as a three-tier production stack. Sol's 95% cache discount drops its effective cost to $0.10 per million cached input tokens, making it the default choice for agentic pipelines with heavy context reuse. The guide covers reasoning effort, async tools, compaction, and delegation as the family's first full operator playbook.
xAI's November GPU Target Doubles to 420,000 -- Colossus Fleet Eyes 1.44 Million by Year End
xAI plans to bring 420,000 Nvidia GPUs online in November, nearly twice the 220,000 figure previously circulated for the same window. Combined Colossus 1 and 2 clusters could reach 1.44 million GPUs by year end. Analysts put 2H 2026 AI CapEx at roughly $53 billion.