AI Updates on 2026-02-18

Anthropic releases Claude Sonnet 4.6 with 1M token context window, improved coding and computer use capabilities, and adjustable effort/thinking modes @claudeai
Alibaba launches Qwen 3.5 Plus on Vercel AI Gateway with 1M token context and adaptive tool use @Alibaba_Qwen
Alibaba releases Qwen3.5-397B-A17B-FP8 weights with SGLang support merged and vLLM support coming @Alibaba_Qwen
Google introduces Lyria 3 music generation model in Gemini, creating 30-second tracks with vocals and lyrics from text or images @GeminiApp
Perplexity adds Claude Sonnet 4.6 for all Pro users and Opus 4.6 option for Max users in browser agent @comet

World Labs raises $1B from AMD, Autodesk, Fidelity, NVIDIA and others to build spatially coherent 3D world generation @theworldlabs
Canva reaches $4B revenue as LLM referral traffic rises, demonstrating AI-driven growth @TechCrunch
Anthropic clarifies Claude Code OAuth policy after confusion, allowing experimentation but requiring API keys for commercial use @trq212
Microsoft reveals Office bug exposed customers' confidential emails to Copilot AI, raising enterprise security concerns @TechCrunch
Coding agents now responsible for 16-23% of GitHub contributions with numbers climbing fast @mikeldking

Anthropic research finds autonomy is co-constructed by model, user and product, requiring post-deployment monitoring beyond pre-deployment evals @AnthropicAI
Analysis shows Claude Code users shift from approving each action to delegating with interruptions as they gain experience @AnthropicAI
73% of agent tool calls on Anthropic API have human in loop, only 0.8% are irreversible actions @AnthropicAI
Google expands SynthID watermarking to audio and adds verification tools in Gemini for identifying AI-generated content @Google

Developer migrates micro-SaaS testimonial service to self-hosted solution in 20 minutes using AI agent, saving $120/year @shanselman
Figma MCP server enables pushing Claude Code prototypes directly to Figma canvas for design iteration @claudeai
Oscar Health migrates 600+ engineers from complex Jira instance to Linear in just over one month @linear
AI Dungeon reduces inference costs from $0.20 to $0.05 per million tokens using NVIDIA Blackwell GPUs and TensorRT LLM @NVIDIAAI

OpenAI introduces EVMbench benchmark measuring AI agents' ability to detect, exploit and patch smart contract vulnerabilities @OpenAI
Microsoft Project Silica publishes Nature paper on encoding data in borosilicate glass for 10,000-year preservation @MSFTResearch
NVIDIA releases FastGen open-source toolkit converting slow diffusion models to few-step generators for real-time AI @ArashVahdat
Toyota Research achieves SOTA using NVIDIA Cosmos-style world models across dynamic view synthesis, teleop augmentation and navigation @NVIDIAAIDev