Live Feed
Anthropic Launches AI Vulnerability Scanner with $104M Coalition — Amazon, Apple, Google, Microsoft, NVIDIA Back Project Glasswing
Claude Code Security entered limited preview on April 11 for Enterprise and Team accounts. It uses AI reasoning to trace data flows and identify vulnerabilities missed by static analysis. The launch is part of Project Glasswing, a $104M industry initiative backed by five major tech companies and 40+ critical infrastructure organisations.
KellyBench: Every Frontier AI Model Loses Money Betting the Premier League
General Reasoning releases KellyBench, a benchmark that forces LLM agents to manage a £100,000 bankroll across a full Premier League season. Every tested model loses money. Claude Opus 4.6 finished best at -11% ROI. Several models went to £0.
NVIDIA Vera Rubin: 50 PFLOPS FP4, 3nm Chiplets, and a New GPU Built Purely for Long-Context Inference
NVIDIA's Vera Rubin architecture lands H2 2026 with 336 billion transistors on TSMC 3nm — a 1.6x density jump from Blackwell — and 50 PFLOPS FP4 per GPU. A new product, the Rubin CPX, targets long-context inference with no Blackwell equivalent.
Anthropic Run-Rate Revenue Hits $30B, Company Explores Custom AI Chips
Anthropic's annualised revenue has surpassed $30 billion — up from $9 billion at end of 2025 — driven by accelerating Claude demand. The company is now exploring designing its own AI chips, mirroring moves by Meta and OpenAI.
ByteDance Launches Seeduplex: First Production-Scale Full-Duplex Voice AI
ByteDance deploys Seeduplex in Doubao, the first voice AI model that listens and responds simultaneously at scale. False interruption rates drop 50% versus the prior half-duplex system. End-of-turn detection is 250ms faster.
Image Generation API Shootout 2026: GPT Image 1.5 Leads Text, Flux 2 Pro Leads Price, Imagen 4 Leads Photo
A comprehensive April 2026 analysis of every major image generation API finds three distinct category leaders — with per-image costs ranging from $0.02 to $0.12 depending on provider, resolution, and quality tier. DALL-E 3 remains most integrated but has lost the quality crown.
Intel and Google Sign Multiyear AI Infrastructure Pact — CPUs Fight Back Against the GPU Monoculture
Intel and Google formalized a multiyear deal on April 9 to deepen Xeon deployment across Google Cloud and co-develop custom Infrastructure Processing Units (IPUs). No financial terms were disclosed, but the partnership signals a deliberate counter-narrative: AI infrastructure needs more than GPUs.
Meta Locks In $21B CoreWeave AI Cloud Through 2032, Anchors Vera Rubin Rollout
CoreWeave and Meta have expanded their existing AI infrastructure agreement to $21 billion through December 2032, with capacity spanning multiple locations and including early Vera Rubin GPU cluster deployments.
OpenAI Launches $100/Month Codex Pro 5X Plan After Token-Based Metering Guts Business Tier
OpenAI introduced a $100/month Codex Pro 5X plan on April 9 delivering 200-1,000 GPT-5.4 messages per 5-hour window. A simultaneous switch to token-based metering has left Business-tier developers reporting effective throughput drops of 2.5-5x, forcing mid-session model swaps.
Anthropic Launches Claude Managed Agents: $0.08/Hr to Run Your AI Agents in the Cloud
Anthropic opened Claude Managed Agents to public beta on April 8, 2026, offering a fully managed runtime for long-running AI agents at $0.08 per session-hour plus standard token costs. Notion, Asana, Sentry, Rakuten, and Atlassian are already in production.
Same Model, 17 More Fixes: Augment Code Leads SWE-Bench Pro by Outscaffolding Claude Code
Augment Code's Auggie CLI scored 51.80% on SWE-bench Pro — 2 points above Cursor and 17 problems above Claude Code — despite all three agents running the same Claude Opus 4.5 model. The entire performance gap came from Augment's semantic codebase indexing engine, not from the underlying model.
DeepSeek R2 Scores 92.7% AIME on 32B Dense Model — Runs on a Single RTX 4090
DeepSeek has abandoned its 671B MoE architecture for a 32-billion-parameter dense transformer that scores 92.7% on AIME 2025, runs on consumer hardware with a single 24 GB GPU, and undercuts Western frontier reasoning APIs by roughly 70% on token cost. MIT license.
Meta Enters the Frontier: Muse Spark Debuts at #4 on the Intelligence Index
Meta Superintelligence Labs launched Muse Spark on April 8 — the first frontier-class model from Meta since Llama 4 and their first proprietary, closed-weights release. It scores 52 on the Artificial Analysis Intelligence Index, placing it 4th behind Gemini 3.1 Pro, GPT-5.4, and Claude Opus 4.6.
OpenAI Publishes Superintelligence Transition Blueprint — Wealth Redistribution, Workforce Protection, Federal Preemption
OpenAI released a landmark policy document in April 2026 outlining how it believes governments should manage the transition to superintelligent AI: universal income mechanisms, workforce safeguards, energy investment, and federal preemption of state AI laws.
AI API Pricing in 2026: A 97% Drop in Three Years and Still Falling
GPT-4 cost $30 per million input tokens when it launched in 2023. Today, comparable-quality models are under $1/M — a 97% reduction in three years. Anthropic averages $6.27/M across its lineup; Meta averages $0.17/M. The price war is far from over.