Skip to content
AIOpen Source

DeepSeek V4 Flash aces ARC-AGI for less than your morning coffee

hackernews / top 1d ago 7

DeepSeek released V4 Flash 0731 on July 31, 2026, offering three reasoning variants (Max, High, Low). At max effort it scores 89.0% on ARC-AGI-1 Semi-Private at $0.02 per task and 61.4% on ARC-AGI-2 Semi-Private at $0.04 per task. The model is available on HuggingFace with an accompanying arXiv paper, and detailed per-task pass/fail breakdowns are provided across 400 ARC-AGI-1 and 120 ARC-AGI-2 public eval tasks.

Read full article →

More AI

Google's AI Brain Trust Pulls a Rolling Stones, Starts New Band

tldr / webdev 1d ago 9

Google lost both Senior Fellows, two DeepMind leaders, and Hassabis stepped back — all in one day; Sundar says it's fine, anyway.

Cloudflare builds a browser for bots, and honestly the bots seem grateful

hackernews / top 1d ago 8

Cloudflare built an entire browser from scratch in 12 weeks so AI agents wouldn't suffer Chromium's memory bloat; your 47 open tabs reportedly unmoved.

Big Tech agrees on a JSON file, AI agents everywhere breathe sigh of relief

tldr / ai 1d ago 8

Vercel, OpenAI, Microsoft, AWS, and GitHub agreed on a shared JSON manifest for AI agent plugins so you package once and run anywhere; world peace still pending, NBD.