Live Feed
DeepSeek V4 Flash Ships at $0.14/M: 79% SWE-Bench Verified on 13B Active Parameters
DeepSeek's efficient V4 variant goes live today with 284B total / 13B active MoE parameters, scoring 79.0% on SWE-Bench Verified at 18x less per token than GPT-5.4. The full V4 family — Pro and Flash — is now on API with 1M standard context and MIT-licensed weights.
DeepSeek-V4-Pro: Open-Source Coding SOTA, 1.6T Params, 1M Context, MIT License
DeepSeek released V4-Pro — 1.6T parameters, 49B activated, 1M context — claiming best open-source model on coding and reasoning. LiveCodeBench 93.5% beats all frontier models. Terminal-Bench 67.9%. Weights are MIT licensed and on HuggingFace now.
Perplexity Research: Post-Training, Not Base Model Choice, Determines Search Quality
Perplexity's published SFT + RL pipeline turns open Qwen models into search tools that match or beat GPT-5.4 on factual benchmarks at lower cost. The core claim is that search quality is a function of how you tune, not what you start with.
GPT-5.5 Leads Intelligence Index at 60 — and Hallucinates at 86%, the Highest Rate at the Frontier
Artificial Analysis' full benchmark sweep of GPT-5.5 reveals a split performance profile: the model sets the highest-ever AA-Omniscience accuracy at 57% while posting an 86% hallucination rate — more than double Claude Opus 4.7 max at 36%. The cost story is more favourable.
80% of Claude's Weekly Users Earn $100K+. Only 37% of Meta AI's Do.
An Epoch AI/Ipsos survey of 5,000 US adults reveals a sharp income stratification in AI adoption. Claude is the most income-concentrated major AI platform by a significant margin. Meta AI is running the opposite strategy entirely.
Anthropic Postmortem: Three Overlapping Claude Code Changes Caused Six Weeks of Quality Degradation
Anthropic published an engineering postmortem tracing Claude Code quality complaints to three separate changes made between March 4 and April 16. All were reversed by April 20. The API was never affected. Usage limits reset for all subscribers on April 23.
OpenAI Launches GPT-5.5: 82.7% Terminal-Bench, 58.6% SWE-Bench Pro, Cheaper Per Codex Task
OpenAI released GPT-5.5, codenamed Spud, to Plus, Pro, Business, and Enterprise users in ChatGPT and Codex. It sets a new Terminal-Bench 2.0 record at 82.7%, matches GPT-5.4 per-token latency, and uses fewer tokens per Codex task. API access pending safety review.
Sony's Ace Robot Beats Elite Table Tennis Players Under Official ITTF Rules — Published in Nature
Sony AI published a Nature cover paper on April 23 describing Ace, the first robot to defeat elite and professional human table tennis players under official tournament rules. Nine active-pixel cameras plus three event-based gaze systems give it 20.2ms end-to-end latency. Trained entirely in simulation via model-free reinforcement learning, it transferred directly to physical competition.
SpaceX's $60B Cursor Option Is Not a Purchase. It Is a Wager on Who Stays Alive.
SpaceX locked in a one-year option to acquire Cursor for $60 billion, or pay $10 billion for the joint work. The number is the headline. The logic underneath it is about something colder: every lab Cursor depends on to run its product is now shipping a direct competitor.
SeedDance 2.0 Holds Both Arena Video #1 Spots at Elo 1450 — 79 Points Ahead of Google Veo 3.1
ByteDance's Dreamina SeedDance 2.0 720p ranks #1 on both Arena's Text-to-Video and Image-to-Video leaderboards, with Elo 1450 and 1449 respectively. It leads Google Veo 3.1 audio 1080p by 79 points on T2V — and does it at 720p.
DeepSeek First External Raise Doubles to $20B in Five Days as Tencent and Alibaba Move In
DeepSeek is in active talks with Tencent and Alibaba at a valuation exceeding $20 billion — double the $10 billion floor it was seeking less than a week ago. The round would be the company's first external funding since its 2023 founding, targeting at least $300 million.
Google Cloud Commits $750M to Accelerate Agentic AI Across Its 120,000-Partner Ecosystem
At Cloud Next in Las Vegas, Google Cloud announced a $750 million fund for its global partner network — covering prototyping, deployment, upskilling, and enterprise agent development. Named enterprise-ready agents from Adobe, Atlassian, Salesforce, Workday and others ship inside Gemini Enterprise on day one.
Google Splits the TPU Line in Two: TPU 8t Trains, TPU 8i Infers
At Google Cloud Next 2026, Google announced its eighth-generation TPUs as two purpose-built chips for the first time. TPU 8t targets training at 3x the compute of the prior generation; TPU 8i targets low-latency inference with SRAM, purpose-built for AI agent workloads.
Japan's FSA Convenes Mythos Risk Meeting With MUFG, SMFG, Mizuho, Bank of Japan, and Tokyo Stock Exchange
Japan's financial regulator is convening an emergency session with the country's five largest financial institutions over Claude Mythos, treating the model as a systemic operational risk alongside liquidity stress and cyberattacks. The specific concern: AI that can scan codebases at machine speed compresses the discovery-to-exploitation window in ways banks cannot patch in real time.
Meta Installs Keystroke and Screen Capture Software on US Employee Computers to Train AI Agents
Internal memos obtained by Reuters reveal Meta is deploying tracking software called Model Capability Initiative (MCI) on US employee workstations — capturing mouse movements, clicks, keystrokes, and periodic screen snapshots — to generate training data for autonomous AI agents that can perform office tasks.