Easy AI Daily — February 18, 2026

A roundup of AI industry news compiled by Easy AI Daily (zhichai.net), covering model releases, agent tooling, infrastructure, research, products, industry moves, and safety.

Models & Capabilities

Anthropic releases Claude Sonnet 4.6: 1M context, near-Opus quality, new cost tradeoffs

Anthropic released Sonnet 4.6, billed as the "most capable Sonnet," 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 achieving that score consumed ~4.8x the tokens of Sonnet 4.5 (and more than Opus 4.6), so total cost on complex tasks may be higher. Reported benchmarks: 79.6% on SWE-bench Verified, 58.3% on ARC-AGI-2. In internal preference tests, users preferred it over Opus 4.5 in 59% of cases. Early reports of function-name hallucinations and degraded structured output were quickly fixed.

Links: Official announcement | Benchmark summary | Artificial Analysis on token usage

Qwen3.5-397B-A17B: Alibaba's open-source 397B MoE targeting GPT-5.2/Opus

Alibaba released Qwen3.5-397B-A17B: 397B total / 17B active MoE parameters, native 262K context extendable to 1M, 200+ languages, and multimodal (text/image/video). Official claims place it near GPT-5.2, Claude Opus 4.5, and Gemini 3 Pro on instruction following, multilingual knowledge, and video understanding benchmarks. Community GGUF quantizations exist for high-memory Macs or multi-GPU local runs, though VRAM demands make it primarily a cloud/cluster model.

Links: Qwen blog | Hugging Face model | Unsloth GGUF

GLM-5 technical report: DSA, asynchronous RL, and agent training

Zhipu published the GLM-5 technical report and demo. Community tests show strong results on benchmarks like WeirdML, approaching or exceeding some closed models. The report highlights the DSA architecture, asynchronous RL infrastructure, and agent-focused RL algorithms. Some developers report weaker performance in certain IDE integrations, likely due to inference-stack differences.

Links: arXiv:2602.15763

Tiny Aya: Cohere's 3.35B multilingual model "runnable on phones"

Cohere Labs released Tiny Aya: 3.35B parameters, 70+ languages, designed for mobile/edge deployment, trained on only 64 GPUs with a detailed technical report. The goal is a practically deployable multilingual assistant rather than top benchmark scores.

Agents & Tooling

• OpenClaw boom and backlash: The multi-agent desktop framework went viral but users report Anthropic account bans for OAuth-based Claude access, deemed "reverse engineering / unauthorized third-party access." Security concerns center on granting LLMs read/write access to machines, browsers, and secrets (prompt injection, malicious skills). Showcase | ClawRecipes
• Harness engineering (LangChain): Turn production traces into automated evals and self-verification loops rather than relying on offline benchmarks; includes TerminalBench and LangSmith Insights scheduled evaluations. OpenAI's related post
• MCP payment signals (SEP-2007): A proposal to let MCP servers signal paid-tool requirements (starting with X402) so agents can spend within budgets. Maintainers prefer lighter URL-based approaches for now. PR #2007
• DirectShell: Uses OS accessibility layers as a universal app interface instead of screenshot+coordinate clicking, effectively turning any windowed app scriptable like the web. GitHub

Infrastructure & Hardware

• GPU MODE kernel engineering: Custom persistent CUDA matmul kernels reaching ~350→368 TFLOPS; FlashInfer MoE baseline achieving 5.7x speedup on B200; discussions on FP8/FP4 numeric error, Nsight Compute vs CUDA events, and approaching theoretical peak FLOPs. matmul_V3 | FlashInfer baseline
• CoDA-GQA-L: Bounded-memory attention fixing KV cache at 384 slots/layer (256 recent, 64 landmark, 64 EMA summary), compressing KV usage from 160GB to 136MB — practical for 1M-context apps. Paper (Zenodo) | Code
• MXFP4/NVFP4 caveat: FP4 benefits require Blackwell (compute capability ≥12.0) native Tensor Cores; emulated FP4 on Ampere-class GPUs is often slower than FP8. Unsloth docs
• Cloud inference outages: OpenRouter mass 401s (war-room response), Perplexity API 401s with remaining balance, HF Inference Endpoints returning 500s with green status pages. Practical lesson: have endpoint-rebuild runbooks for production.

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 compressible references; 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 — cold water on "stack agents, get AGI" narratives.
• Every Eval Ever (EvalEval consortium): A BIDS-like unified data/metadata standard for LLM benchmarks, motivated by issues like Cybench's non-randomized flags inflating scores. Project page
• Preventative steering: Adding steering vectors to intermediate representations to test whether models still hit targets; proposed as representation-space data augmentation. arXiv:2602.15029

Products & Applications

• Sonnet 4.6 rolls out widely: Available within a day in Cursor, Windsurf, Cline, Perplexity, OpenRouter. Cursor finds it notably better than 4.5 on long tasks but slightly below Opus 4.6 in "intelligence"; Cline's internal A/B showed 70% developer preference for 4.6. Emerging pattern: Sonnet 4.6 as the high-value workhorse, Opus for maximum quality.
• PolyAI: $200M raise (Nvidia, Khosla participating) plus Agent Studio Lite, which generates a working voice agent from a URL in 5 minutes; 3-month free trial.
• SpeakType: Fully local, open-source Mac dictation app built on Whisper; community questions focus on RAM/CPU usage, VAD preprocessing, and comparisons to tools like Handy. GitHub
• Manus Agents: A personal agent with long-term memory and tool integrations (Gmail, Notion) delivered via Telegram, generating videos, decks, and web pages.
• Food Truck Bench: 12 LLMs each given $2,000 and a food truck for a 30-day simulation; only 4 survived. Claude Opus 4.6 earned ~$49K, GPT-5.2 ~$28K; nearly every model that took loans went bankrupt. Combined with Vending-Bench 2: aggressively expanding/borrowing agents tend to die in realistic settings. Reddit thread

Industry & Business

• Mistral acquires Koyeb: Folding the serverless platform and team into Mistral Compute, shifting from API-only model vendor toward a model+compute cloud company. Announcement
• Grok 4.20 skepticism: Marketed as new architecture, but logs show model ID grok-4-1-thinking-1129, mocked as "four Grok 4.1 agents in a trench coat"; also criticized for political bias and underperformance versus peers.
• Anthropic military policy: Claude permitted for limited US military use with two explicit red lines: no mass surveillance and no autonomous weapons.
• Waymo economics: 6th-gen robotaxi platform at ~$70K per vehicle, projected to halve by 2028; over 500K weekly driverless rides, 3x annual growth. Analysis

Policy, Governance & Safety

• OpenClaw OAuth bans: Anthropic's TOS treats third-party desktop-control OAuth access as "reverse engineering / unauthorized software" — a warning to desktop-agent teams to use official APIs or self-hosted models.
• Dark-web overlap: BASI community reports of GitLab maintainer-access projects auctioned on the dark web ($200 start, $2,000 flash sale) and Tor-hosted "unrestricted AI" services potentially laced with malware or Claude abuse — LLM tech and gray/black markets now intersect.
• API key leakage: An OpenRouter user's key leaked despite gitignore, drained $10 in 20 minutes via a third-party environment. Reminders: least privilege, key rotation, full-repo secret scans before launch.
• Cybench flag flaw: Stanford's security benchmark used fixed flags from well-known CTFs, letting models recall them from training data and inflate success rates; randomizing flags caused sharp drops — a caution for security eval design. Cybench

Source: Easy AI Daily (zhichai.net)