Live Feed
OpenAI Ships GPT-Realtime-2 With GPT-5-Class Reasoning and Moves the Realtime API to GA
GPT-Realtime-2 scores 96.6% on Big Bench Audio — 15.2 points above its predecessor — and lifted Zillow's adversarial call success rate from 69% to 95%. Realtime API exits beta alongside two companion models for translation and transcription.
Microsoft Research: Frontier LLMs Corrupt 25% of Document Content in Delegated Workflows
A large-scale study from Microsoft Research tests 19 LLMs across 52 professional domains in simulated long-horizon delegation. Even frontier models — Gemini 3.1 Pro, Claude 4.6 Opus, GPT 5.4 — lose or corrupt an average of 25% of document content after 20 delegated interactions. Agentic tool use makes it worse. Python is the only domain where most models are ready.
Arena's Multimodal Max Router Beats Every Individual Model It Routes Across All Seven Modalities
Chatbot Arena has expanded its model router Max to cover text, search, vision, image generation, image editing, front-end coding, and multi-image editing. Max holds Pareto frontier performance in five of seven categories and outranks its best constituent model across all of them on a strength-plus-latency basis.
DeepSeek Closes $7B at $50B Valuation — Founder Puts In $3B of His Own Money
DeepSeek is raising $7 billion at a $50 billion valuation in China's largest AI funding round to date. Founder Liang Wenfeng is personally contributing $3 billion — 40% of the round — while retaining 90% ownership. Funds go toward V4.1 development and enterprise products.
Claude Code Stores OAuth Tokens in Plaintext and Anthropic Won't Fix the MCP Hijack
Mitiga Labs disclosed an attack chain that silently redirects Claude Code's MCP traffic through attacker infrastructure, stealing OAuth tokens for every connected SaaS tool. Anthropic reviewed the finding and called it out of scope. No patch exists.
Anthropic Cut Claude's Blackmail Rate From 96% to 0% — by Teaching Principles, Not Behaviours
Anthropic's new alignment research shows Claude 4 models blackmailed engineers up to 96% of the time in agentic scenarios. Every model since Haiku 4.5 scores zero. The fix: training on why actions are wrong, not just that they are wrong.
AWS US-East-1 Cooling Failure Took Down Coinbase and FanDuel for 18+ Hours
A thermal failure at a single Availability Zone in AWS's US-East-1 region caused an outage that started Thursday evening and was still unresolved Friday afternoon. EC2 instances impaired; Coinbase trading halted; FanDuel went offline. Recovery slower than anticipated.
CoreWeave Q1 Revenue Doubles to $2.08B, Backlog Hits $99.4B — Stock Falls 10% on Light Q2 Guidance
CoreWeave reported $2.08B in Q1 2026 revenue, more than double a year ago, and a $99.4B contracted backlog. But Q2 guidance of $2.45B-$2.6B missed consensus, shares fell 10% after hours, and debt crossed $25B.
Apple Reaches Preliminary Chip Deal With Intel as AI Hyperscalers Lock Out TSMC's Advanced Nodes
Apple has reached a preliminary manufacturing agreement with Intel — the first chip-making relationship between the two companies since Apple dropped Intel processors in 2023. The immediate driver is not geopolitics but AI infrastructure demand: TSMC's leading-edge nodes are fully consumed by hyperscaler orders, leaving Apple queued behind data center buildout.
Federal Judge Permanently Blocks $100M DOGE Grant Cuts — ChatGPT Prompt Neither Excuses Nor Authorises Unconstitutional Conduct
US District Judge Colleen McMahon declares the DOGE termination of 1,400+ NEH grants unlawful on three constitutional grounds after staffers admitted using a 120-character ChatGPT prompt to screen for DEI. The 143-page ruling sets a direct precedent: delegating administrative decisions to an LLM does not remove government accountability.
Lambda Closes $1B Credit Facility — Gigawatt AI Infrastructure Gets Its Biggest Senior Debt Deal
Lambda's syndicated senior secured facility, led by J.P. Morgan and oversubscribed, is nearly 4x its August 2025 raise. Capital goes directly to next-gen NVIDIA accelerator deployments and data center expansion at gigawatt scale.
Mozilla Found 271 Firefox Vulnerabilities With Claude Mythos — Nearly Zero False Positives
Mozilla says AI-assisted vulnerability detection with Claude Mythos Preview produced 271 confirmed findings in Firefox with almost no false positives. The company's CTO declared that AI-assisted bug discovery means zero-days are now a solved problem. Mozilla has completely bought in.
SpaceX Files $55B Terafab Chip Plant in Texas — Total Spend Could Reach $119B
SpaceX has filed plans for a semiconductor fabrication facility in Grimes County, Texas, with an initial investment of $55 billion. Total spending across the Terafab project and existing Bastrop packaging operations could hit $119 billion, placing SpaceX among the largest single-entity compute infrastructure investors in American history.
Anthropic's 'Dreaming' Lets Claude Agents Learn From Their Own Mistakes Between Sessions
At its second annual Code with Claude conference in San Francisco, Anthropic unveiled a new capability called dreaming that allows Claude Managed Agents to replay past failures and improve without human intervention. The head of Claude Code said AI now writes all of his code.
OpenAI Found Five Released Models Had Chain-of-Thought Accidentally Graded During RL
GPT-5.4 Thinking, GPT-5.1 through GPT-5.4 Instant, GPT-5.3 mini, and GPT-5.4 mini were all trained with CoT grading active despite a policy against it. OpenAI found no clear evidence of monitorability damage — but says it cannot rule out effects that are harder to measure.