English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Easy AI Daily News Digest – February 18, 2026: Claude Sonnet 4.6, Qwen3.5-397B, GLM-5, and More

Forum topic · 小凯 · 2026-03-27

Summary

Easy AI Daily for February 18, 2026 covers major AI model releases and industry developments. Anthropic launched Claude Sonnet 4.6 with 1M token context, topping GDPval-AA at ELO 1633 but consuming roughly 4.8x the tokens of Sonnet 4.5, raising cost concerns. Alibaba open-sourced Qwen3.5-397B-A17B, a 397B-parameter MoE model with 17B active parameters, native 262K context (expandable to 1M), and 200+ language multimodal support. Zhipu published the GLM-5 technical report highlighting DSA architecture and asynchronous RL, while Cohere Labs released Tiny Aya, a 3.35B multilingual model for edge devices. On the agent side, OpenClaw's popularity triggered Anthropic account bans over OAuth usage, MCP discussed payment signaling in SEP-2007, and DirectShell proposed accessibility-layer control to replace screenshot-based agents. Infrastructure news includes CoDA-GQA-L compressing KV cache from 160GB to 136MB, FP4 formats showing gains only on Blackwell GPUs, and cloud outages at OpenRouter and Hugging Face. Industry moves: Mistral acquired Koyeb, xAI released Grok 4.20 amid skepticism, PolyAI raised $200M, and Waymo reported 500K+ weekly driverless rides.

Easy AI Daily News Digest – February 18, 2026

A structured summary of AI industry news for February 18, 2026, covering models, agents, infrastructure, research, products, industry moves, and safety.

Models & Capabilities

  • Anthropic Claude Sonnet 4.6: Touted as the strongest Sonnet yet, with upgrades in coding, computer use, long-context reasoning, and agent planning, plus 1M token context (beta) at the same price as 4.5. It ranks first on third-party GDPval-AA (ELO 1633), but consumed ~4.8x the tokens of Sonnet 4.5 to achieve that score—more than Opus 4.6—so total cost on complex tasks may be higher. Reported benchmarks: SWE-bench Verified 79.6%, ARC-AGI-2 58.3%; users preferred it over Opus 4.5 in 59% of internal preference tests. Early reports of function-name hallucinations and degraded structured output were quickly fixed.
  • Official announcement | Benchmark data roundup | Artificial Analysis on GDPval-AA and token usage
  • Alibaba Qwen3.5-397B-A17B: Open-source MoE with 397B total / 17B active parameters, native 262K context extendable to 1M, 200+ languages, and image/video multimodality. Claims near-GPT-5.2, Claude Opus 4.5, and Gemini 3 Pro performance on instruction following, multilingual knowledge, and video understanding. GGUF quantizations exist, but VRAM demands make it mostly a cloud/cluster model.
  • Qwen blog | Hugging Face model | Unsloth GGUF
  • Zhipu GLM-5 technical report (arXiv:2602.15763): Highlights DSA architecture, asynchronous RL infrastructure, and agent-focused RL algorithms. Strong community benchmark results (e.g., WeirdML), though some developers report weaker performance inside certain IDE integrations.
  • Cohere Labs Tiny Aya: 3.35B-parameter, 70+ language model designed for phones/edge devices; trained on just 64 GPUs with a detailed technical report. Prioritizes deployability over leaderboard scores.
  • Agents & Tooling

  • OpenClaw controversy: The multi-agent framework for controlling real computers went viral, but users connecting via Claude OAuth were reportedly banned by Anthropic for violating terms against unauthorized third-party access. Security concerns include prompt injection and malicious skills given local file, browser, and key access.
  • Harness Engineering: LangChain promotes turning production traces into automated eval loops (with TerminalBench and LangSmith Insights); OpenAI published a related blog post that some criticized as marketing-heavy.
  • MCP payment signaling: SEP-2007 proposes letting MCP servers return a "payment required" signal (starting with X402). Maintainers are cautious, preferring URL redirects for now; not a near-term core priority. See PR #2007.
  • DirectShell: Uses the OS accessibility layer as a universal app interface instead of screenshots and coordinate clicks, aiming to make any desktop app scriptable like the web. GitHub repo.
  • Infrastructure & Hardware

  • GPU MODE community results: Custom persistent CUDA kernels reached ~350→368 TFLOPS; FlashInfer MoE baseline achieved 5.7x speedup on B200. Ongoing discussions on FP8/FP4 numerical error and correct use of Nsight Compute and CUDA events.
  • CoDA-GQA-L: Fixed-size KV cache per layer (384 slots: 256 recent tokens, 64 landmark tokens, 64 EMA summaries) compresses KV cache from 160GB to 136MB on long sequences. Paper on Zenodo | Code.
  • MXFP4/NVFP4 caveats: FP4 speedups require native Blackwell Tensor Cores (compute capability ≥12.0); on Ampere and older cards, emulation is often slower than FP8.
  • Cloud inference outages: OpenRouter suffered widespread 401s; Perplexity API users hit 401s with positive balances; Hugging Face Inference Endpoints returned intermittent 500s despite green status pages. Practical lesson: have scripts ready to fully rebuild endpoints and migrate traffic.
  • Research & Methods

  • Agent World Model: 1,000 executable environments, 35,062 tools, 10K tasks for RL training of tool-using agents.
  • Lossless Context Management (LCM/Volt): Hierarchical DAG-based recoverable-reference compression, reportedly more stable than Claude Code from 32K to 1M context on OOLONG.
  • Moltbook social simulation: 2.6M LLM agents produced 300K posts and 1.8M comments; macro-level culture converged and stabilized, but individual agent influence was statistically negligible.
  • Every Eval Ever: The EvalEval alliance proposes a unified data/metadata standard for LLM benchmarks, akin to BIDS in neuroimaging—motivated by issues like Cybench's unrandomized flags inflating scores.
  • Preventative steering: Discussion of applying steering vectors to intermediate representations to test whether models retain goals under perturbation, potentially as representation-space data augmentation (arXiv:2602.15029).
  • Products & Applications

  • Sonnet 4.6 rollout: Available within a day in Cursor, Windsurf, Cline (free trial; 70% of developers preferred it in internal A/B), Perplexity Pro/Max (Deep Research and browser agent), and OpenRouter. Sonnet 4.6 is becoming the value workhorse; Opus reserved for maximum quality.
  • PolyAI: Raised $200M (Nvidia, Khosla participating) and launched Agent Studio Lite, which generates a working voice agent from a website URL in ~5 minutes; 3-month free trial for brands.
  • SpeakType: Fully local, open-source Mac dictation app based on Whisper; community questions cover RAM/CPU usage, VAD preprocessing, and differences vs. tools like Handy. GitHub.
  • Manus Agents: A personal agent with long-term memory and tool integrations (Gmail, Notion; video/PPT/web generation) accessible via Telegram.
  • Food Truck Bench: 12 LLMs each given $2,000 and a simulated food truck for 30 days; only 4 survived. Claude Opus 4.6 earned ~$49K, GPT-5.2 ~$28K; nearly all models that took loans went bankrupt—echoing Vending-Bench 2 findings that aggressive expansion kills agents.
  • Industry & Policy

  • Mistral to acquire Koyeb: Absorbing the serverless platform into Mistral Compute, shifting from model vendor to integrated model-plus-compute cloud provider.
  • Grok 4.20 skepticism: Community found the model ID still reads grok-4-1-thinking-1129, mocking it as "four Grok 4.1 agents in a trench coat"; criticized for echoing Musk's positions on gender/political topics.
  • Anthropic and the US military: Claude permitted for limited military use with two explicit red lines—no mass surveillance and no autonomous weapons.
  • Waymo economics: Sixth-gen robotaxi platform costs ~$70K per vehicle (projected to halve by 2028), with 500K+ weekly driverless rides growing ~3x annually.
  • Security & Governance

  • OpenClaw bans: Anthropic treats OAuth-based third-party desktop control tools as unauthorized access under its ToS—teams building desktop agents should use official APIs or self-hosted models.
  • Dark-web incidents: Three GitLab projects with maintainer access auctioned starting at $200; "unrestricted AI" services on Tor raise malware and abuse concerns.
  • API key leaks: An OpenRouter user's key leaked despite being in a gitignore file and was drained of $10 in 20 minutes. Advice: least privilege, regular rotation, and full repo scans before deployment.
  • Cybench flaw: Early versions used fixed flags from well-known CTFs, letting models recall answers and inflate success rates; randomizing flags sharply dropped scores—showing how benchmark design flaws can distort security evaluations. Cybench.
*Source: Easy AI Daily (zhichai.net digest)*

Tags

#ai-news#claude-sonnet-4-6#qwen3-5#glm-5#llm-benchmarks#ai-agents#gpu-optimization#ai-policy

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177169259