Latest news from the AI world
Week of March 3–5, 2026: OpenAI's GPT-5.4 launches native computer use and reaches 75% on OSWorld, NotebookLM generates immersive videos with Gemini as director, Codex arrives on Windows, and GitHub enables Copilot Memory by default.
March 3, 2026: Claude Code launches Voice Mode (push-to-talk, free), GPT-5.3 Instant replaces GPT-5.2 in ChatGPT with hallucinations down 26.8%, Gemini 3.1 Flash-Lite in preview. Teaser GPT-5.4, FLUX.2 [pro] 2x faster, ElevenLabs at MWC.
March 2, 2026: Claude Code v2.1.63 delivers /simplify, HTTP hooks and 8 memory fixes, OpenAI signs a classified AI agreement with the Pentagon, Anthropic refuses mass surveillance and threatens to sue. Also: Claude Memory on the free plan, Qwen 3.5 Small Series, Gemini CLI v0.31.0, and deprecations of Gemini 3 Pro and GPT-5.1 in Copilot.
February 27, 2026: OpenAI announces a $110 billion round with SoftBank, NVIDIA and Amazon, an exclusive AWS cloud partnership, Gemini Drop with Lyria 3 and Nano Banana 2, and GitHub launches Agentic Workflows in Markdown.
26 February 2026: Dario Amodei publishes a statement on Anthropic's limits vis‑à‑vis the U.S. Department of War, Perplexity obtains OS-level access on the Samsung Galaxy S26, Claude and Codex go GA for Copilot Business and Pro, NVIDIA unveils LillyPod with 1,016 Blackwell Ultra GPUs.
February 25, 2026: Anthropic buys Vercept to accelerate computer use (Claude Sonnet 4.6 at 72.5% on OSWorld), Perplexity launches Computer — an autonomous multi-model orchestrator, GitHub Copilot CLI goes generally available with MCP and agents support.
February 24, 2026: Anthropic publishes the third version of its Responsible Scaling Policy with a public Frontier Safety Roadmap and quarterly Risk Reports; GPT-5.3-Codex moves to general availability via the Responses API; Meta signs a multi-year AMD deal for ~6 GW of GPU capacity; Qwen 3.5 Medium launches 4 MoE models including a 35B that outperforms the 235B giant; Claude Code v2.1.51 introduces Remote Control from mobile.
February 23, 2026: Anthropic reveals that DeepSeek, Moonshot AI and MiniMax extracted 16 million exchanges via 24,000 fraudulent accounts to distill Claude; OpenAI drops SWE-bench Verified (59.4% defective tests) and recommends SWE-bench Pro; gpt-realtime-1.5, WebSockets in the Responses API, Veo 3.1 templates and Pika AI Selves.
Claude Code v2.1.49-2.1.50 extends multi-worktree agents and fixes 6 memory leaks, GitHub Copilot introduces cross-session persistent memory, and ElevenLabs cuts TTS latency by 40% on February 22, 2026.
Anthropic launches Claude Code Security in preview (500+ vulnerabilities detected in open source), Runway becomes a multi-model hub, and GitHub Copilot extends its model picker to Business/Enterprise on February 20, 2026.
Google launches Gemini 3.1 Pro with an ARC-AGI-2 score of 77.1%, OpenAI expands to India with Tata Group and Stargate infrastructure, Anthropic opens Claude in PowerPoint to Pro plans, and Mistral publishes its Voxtral Realtime technical report.
February 16-18, 2026: Anthropic launches Claude Sonnet 4.6 with improved computer use and 1M token window. Qwen releases Qwen3.5-397B in open-weight Apache 2.0. Google integrates Lyria 3 into Gemini. OpenAI unveils EVMbench, Cohere launches Tiny Aya, Runway opens its Gen-4.5 API.