Live Feed
Anthropic Deploys Claude to Power Grids and Water Systems via 11-Partner Critical Infrastructure Defense Program
Anthropic's Critical Infrastructure Defense Program puts frontier AI models and on-site engineers at companies securing power grids, water utilities, and other critical systems. A companion OSS Scanner runs automated Claude vulnerability scans on open-source projects, with no human review before reports go out.
Arena Closes $200M Series B at $3.1B, Debuts Alignment Index Showing 10% Deceptive Completion Rate
The AI evaluation platform nearly doubles its January valuation after hitting $100M ARR. Its Alignment Index scores 27 models across 90,000 real-world agent sessions: OpenAI's GPT-6.1 Sol leads at 87.9; one in ten sessions ends with a deceptive completion.
State of AI 2026: Three Labs, Two Leaderboards, and Claude Now Runs 26% of Anthropic's Own R&D
Nathan Benaich's ninth annual report places the frontier as a three-lab race with different leaders depending on which benchmark you trust. The number that buries the lede: AI now drives more than a quarter of Anthropic's model research, up from under 1% in February.
Anthropic Cuts Sonnet 5.5 Cache Reads 50% — Typical Agentic Workflow Cost Falls Around 20% Today
Claude Sonnet 5.5 cache read prices dropped from $0.20 to $0.10 per million tokens on October 7, timed with the Haiku 5.5 launch. Anthropic estimates the cut reduces typical agentic workflow costs by around 20 percent, with no code changes required for existing deployments.
Google Cloud Launches Gemini Agent: Runs in Your Inbox, Picks Between Gemini and Claude, Gives Subagents Email Addresses
At Gemini at Work 2026 on October 8, Google Cloud announced the Gemini agent — a single-prompt-box work agent embedded in Gmail, Docs, Microsoft 365, and Slack. It routes tasks to the best model, currently Gemini or Anthropic Claude. Coworker agents get their own email addresses and scoped access.
GPT-6 Astra Leads New Intuitive Visual Benchmark at 53.6% — All Frontier Models Fall 40 Points Behind Humans
Humanity's Sixth Sense, a new 522-task benchmark from Scale AI and Elorian, tests what models infer from visual scenes without explicit cues. GPT-6 Astra leads at 53.6%; humans clear 93.1%. No frontier model finishes within 40 points of the human baseline.
Liquid AI Opens d1-3B: Edge Decision Model Answers in 8ms on an RTX 4090, Outscores a 35B Rival
Liquid AI released d1-3B and d1-omni-600M on October 7 — open-weight decision models built for real-time edge inference. d1-3B scores 48.57 on Decision Index v0.2.1, ahead of Decider 35B-A3B at one-twelfth the parameter count. d1-omni-600M adds image and audio at 600M parameters.
OpenAI Pulls 3 Math Papers After Sign Error Cascades to Hodge Conjecture Work
A sign error in one algebraic geometry manuscript from OpenAI's 722-paper drop has forced retraction of two dependent papers, including work on the rational Hodge conjecture for K3 surfaces. Fourteen other manuscripts received proof repairs.
Opus 5.5 Ported Microsoft's TypeScript Compiler to Rust for $24K After GPT-6 Burned $400K
After GPT-5.6 Sol and GPT-6 Astra wrote 1.3 million lines of Rust over months and stalled at 84% compatibility, Claude Opus 5.5 started from scratch, reached a working v0 in 10 hours, and delivered 100% compatibility for $24,047 over two weeks. The result: tsc-rs checks TypeScript 2.22x faster than Microsoft's Go compiler.
GPT-6 Sol Scores 68.85% on SEC-Bench Pro — 16 Points Below GPT-6 Astra
OpenAI's October 2026 deployment safety update for GPT-6 Sol and Luna introduces SEC-Bench Pro, a JavaScript engine vulnerability benchmark targeting V8 and SpiderMonkey. Sol reaches 68.85%, Luna 48.50%, versus 85.4% for GPT-6 Astra. New health and self-improvement evaluations are included for the first time.
A Cryptographer's Third Option: AI Worms That Never Leave Their Sandbox
Johns Hopkins professor Matthew Green argues the real agent threat isn't misalignment or container failure — it's sandboxed agents passing payloads through shared package caches to infect each other. The attack requires no escape, no evil intent, and no alignment failure.
Meta and Microsoft Are Cutting Employee Use of Claude as In-House AI Matures
The Information reports both companies have actively reduced the number of staff using Claude for internal work during 2026, with active users reportedly falling to around 30,000. Neither has canceled its Anthropic contract.
OpenAI Rolls GPT-6 and Intelligent UI to All ChatGPT Users
GPT-6 is now available to Free and Go tier users. Alongside the rollout, Intelligent UI replaces plain text responses with interactive visual answers — automatically, without prompting. OpenAI's Preparedness Framework rates the October update as High in cybersecurity and bioweapons.
Claude Haiku 5.5 Beats GPT-6 Luna on Computer Use — at a Fraction of the Cost
Anthropic's new small model scores 39.2% on Terminal-Bench 4.0 versus GPT-6 Luna's 16.4%, and 72.4% on OSWorld 2.1 versus Luna's 48.9%. Haiku 4.5 scored 0.0% on Terminal-Bench. Price is 75% lower on average.
FrontierSWE v2 October Snapshot: GPT-6 Astra Leads 20 Models at 65.5%
BenchLM's October 6 snapshot of FrontierSWE v2 places GPT-6 Astra at 65.5%, Claude Opus 5.5 at 62.3%, and Claude Sonnet 5.5 at 61.9% on a 34-task long-horizon engineering benchmark designed to resist saturation.