AI Updates on 2026-02-12

Google releases upgraded Gemini 3 Deep Think achieving 84.6% on ARC-AGI-2 and 3455 Elo on Codeforces, ranking #8 globally among competitive programmers @GoogleDeepMind
Gemini 3 Deep Think earns gold medal-level results on 2025 Physics and Chemistry Olympiads and sets new standard on Humanity's Last Exam with 48.4% without tools @sundarpichai
OpenAI launches GPT-5.3-Codex-Spark in research preview delivering over 1000 tokens per second for Pro users @sama
Gemini 3 Deep Think now converts sketches into 3D-printable STL files by analyzing drawings and modeling complex shapes @GeminiApp
Alibaba demonstrates Qwen3-ASR-0.6B powering on-device real-time transcription on iPhone 15 Pro Max @Alibaba_Qwen

Anthropic raises $30B at $380B valuation with run-rate revenue of $14B, growing 10x annually for three consecutive years @AnthropicAI
Claude Code reaches $2.5B run-rate revenue in under one year, doubling weekly active users since January 1 @AnthropicAI
NVIDIA rolls out Codex company-wide to 30,000 engineers with cloud-managed admin controls and US-only processing @OpenAIDevs
Spotify reports best developers haven't written code since December due to AI productivity gains @TechCrunch
OpenRouter weekly token consumption explodes 12.7x to 12.1 trillion tokens per week, matching Azure's entire inference volume @deedydas
AI inference startup Modal Labs in talks to raise at $2.5B valuation @TechCrunch

Anthropic commits to covering 100% of electricity price increases and grid upgrade costs from data centers to protect ratepayers @AnthropicAI
Anthropic contributes $20M to Public First Action, bipartisan organization focused on AI policy mobilization @AnthropicAI
AI Now Institute releases interventions on open source, digital sovereignty, linguistic diversity, and multilateralism in AI governance @AINowInstitute

Duke University's Wang Lab uses Gemini 3 Deep Think to optimize growth of complex crystals for high-temperature semiconductors @GoogleDeepMind
Perplexity launches Memory feature working with Model Council, enabling multiple frontier reasoning models to collaborate on user data @AravSrinivas
Andrew Ng releases course on A2A protocol standardizing agent communication, built with Google Cloud and IBM Research @AndrewYNg

Research demonstrates GPT-5 mini outperforms highly trained graduate students at information retrieval tasks at 1000x lower cost @emollick
Google DeepMind introduces Aletheia agent based on Gemini Deep Think for PhD-level mathematics, autonomously generating papers and solving open Erdős conjectures @quocleix
Karpathy releases microGPT: full GPT training and inference in 200 lines of pure dependency-free Python @karpathy
Jeff Clune's team develops meta-learning system where AI agents automatically design better memory mechanisms for themselves @jeffclune