AI Updates on 2026-02-07

Anthropic releases Claude Opus 4.6 in fast mode, running 2.5x faster than standard version, now available via Claude Code and API @claudeai
Anthropic grants all Claude Pro and Max users $50 in free extra usage credits for fast mode Opus 4.6 in Claude Code @_catwu
Cursor integrates Opus 4.6 fast mode at $30 input/$150 output per million tokens, offering 50% discount for 10 days @cursor_ai
Google updates Veo 3.1 with portrait mode support, more expressive movement control, and state-of-the-art upscaling to 4K @JeffDean

Perplexity launches Model Council feature enabling parallel research across GPT-5.2, Claude Opus 4.6, and Gemini 3 Pro with consensus analysis @AravSrinivas
NVIDIA reports Cursor helping ship 3x more committed code across large codebases by accelerating onboarding and automating workflows @NVIDIAAI
Heroku transitions to sustaining engineering model, ending new Enterprise contracts while focusing investments on enterprise-grade AI deployment @GergelyOrosz
Benchmark raises $225M in special funds to double down on Cerebras investment @TechCrunch
X API launches pay-per-use pricing with 20% cashback in xAI credits for developers spending on X API @xai

Waymo uses Google's Genie 3 world model to generate photorealistic interactive simulations of rare driving events for autonomous vehicle training @sundarpichai
Strong DM launches "Software Factory" approach where code is neither written nor reviewed by humans, spending $1,000/engineer/day in tokens @simonw
Apple working to integrate AI chatbots like ChatGPT into CarPlay for in-vehicle assistance @TechCrunch
Anthropic adds WordPress integration to Claude, enabling easier site monitoring and management @TechCrunch

EchoJEPA foundation model trained on 18M heart ultrasound videos reduces error on cardiac function metrics by 20% using JEPA architecture @ylecun
PyTorch releases fused Triton kernel for Mamba-2 achieving 1.5x-2.5x speedups on NVIDIA A100 and H100 GPUs @PyTorch
Berkeley AI research reveals LLMs can embed hidden instructions through data subsets without system prompts or visible signals @berkeley_ai
MIT physicists demonstrate new form of magnetism potentially enabling faster, denser, lower-power spintronic memory chips @MIT