Skip to content
Source · Daily Brief

AI Daily Brief — 17 February 2026

A four-launch Tuesday. Anthropic ships Claude Sonnet 4.6 — twelve days after Opus 4.6 — with full upgrades across coding, computer use, long-context reasoning, agent planning and design. NIST‘s newly stood-up Center for AI Standards and Innovation (CAISI) launches the AI Agent Standards Initiative, explicitly naming the Model Context Protocol ecosystem as a community pillar. xAI ships Grok 4.20 Beta with weekly rapid-learning updates and a four-agent collaboration architecture. Meta and NVIDIA announce a roughly $50B multi-year, multi-generational AI infrastructure deal — millions of Blackwell + Rubin GPUs plus millions of Grace / Vera CPUs, making Meta the first Big Tech firm to commit to standalone NVIDIA CPUs for inference. Cursor 2.5 ships a Plugin Marketplace and async subagents.

Top stories

  • Claude Sonnet 4.6. Scores 79.6% on SWE-bench Verified — just 1.2 points behind Opus 4.6 — 72.5% on OSWorld, and 59.1% on Terminal-Bench 2.0. Pricing held at $3 / $15 per M tokens (5× cheaper than Opus). 1M-token context window available in API beta. Becomes the default model on claude.ai and in Claude Cowork for Free and Pro users. via Anthropic
  • NIST — AI Agent Standards Initiative. CAISI organises the initiative around three pillars: industry-led agent interoperability and security standards (with US leadership in international bodies); community-led open-source agent protocols including the Model Context Protocol ecosystem; and fundamental research into agent authentication and identity infrastructure. RFI on Agent Security due 9 March; ITL Concept Paper on Agent Identity and Authorization due 2 April; sector-specific listening sessions begin in April. via NIST
  • xAI — Grok 4.20 Beta. Weekly rapid-learning model updates based on real-world feedback, medical-document analysis via photo upload, improved engineering reasoning, and a four-agent collaboration architecture: Grok coordinates while Harper handles research, Benjamin logic/math, and Lucas contrarian analysis. Free on grok.com with usage limits; unlimited via SuperGrok at $30/month. API access scheduled for 10 March. via xAI
  • Meta × NVIDIA — ~$50B multi-generational deal. Meta commits to millions of NVIDIA Blackwell and Rubin GPUs plus millions of Grace/Vera CPUs — the first Big Tech standalone CPU commitment to NVIDIA, for inference rather than training. Bundle includes Spectrum-X Ethernet networking and Confidential Computing for WhatsApp. Estimated value ~$50B; supports Meta’s $115-135B 2026 capex guidance. Zuckerberg frames it as “crucial to delivering personal superintelligence to everyone in the world” — raising fresh questions about Meta’s in-house MTIA chip program. via NVIDIA
  • Cursor 2.5 — Plugin Marketplace + async subagents. Launch partners Amplitude, AWS, Figma, Linear and Stripe. Plugins bundle skills, subagents, MCP servers, hooks and rules into a single install. Subagents can now run asynchronously and spawn their own subagents (tree-of-work). Sandbox gains granular network-access controls and per-directory/per-file filesystem controls; lower-latency streaming for parallel subagent execution. via Cursor

Markets & money

Meta shares move on the NVIDIA announcement; SiliconANGLE flags investor questions about whether the scale of the commitment signals delays or scope reductions to Meta’s custom MTIA roadmap. TechCrunch publishes its February-17 roundup of 17 US-based AI companies that have crossed the $100M raise threshold so far in 2026 — context for a record-setting funding environment in which roughly 83% of February’s $189B global venture total concentrates in three megadeals (OpenAI $110B, Anthropic $30B, plus one other). via TechCrunch

Cadence note

Developers in early access prefer Sonnet 4.6 over Opus 4.5 in 59% of head-to-heads, citing better instruction following and less over-engineering. With Sonnet 4.6, Grok 4.20 Beta and Cursor 2.5 all on one Tuesday, February tracks toward seven frontier-model releases in a single month across Google, Anthropic, OpenAI, xAI and Alibaba.

Quiet corners

No first-party shipping from OpenAI, Google DeepMind or Microsoft AI. Gemini 3.1 Pro is queued for Thursday.

By the numbers

  • Sonnet 4.6: SWE-Bench Verified 79.6% · OSWorld 72.5% · Terminal-Bench 2.0 59.1% · $3 / $15 per M tokens (5× cheaper than Opus)
  • Grok 4.20 Beta: weekly retraining cadence · 4-agent collaboration (Grok · Harper · Benjamin · Lucas) · SuperGrok $30/mo
  • Meta × NVIDIA: ~$50B deal · millions of Blackwell + Rubin GPUs · millions of Grace + Vera CPUs · Meta 2026 capex $115-135B
  • Cursor 2.5 Marketplace partners: Amplitude · AWS · Figma · Linear · Stripe
  • Sonnet 4.6 dev preference vs Opus 4.5: 59% in head-to-head

Compiled by AI Feed’s editor from publicly reported announcements on 17 February 2026.