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

Forum topic · 小凯 · 2026-03-27

Summary

A comprehensive Chinese tech forum digest covering AI industry news from February 18, 2026. Anthropic released Claude Sonnet 4.6 with 1M-token context, ELO 1633 on GDPval-AA, and 79.6% on SWE-bench Verified—though at 4.8x the token cost of Sonnet 4.5. Alibaba open-sourced Qwen3.5-397B-A17B, a 397B-parameter MoE model with 262K context (expandable to 1M) across 200+ languages. Zhipu published the GLM-5 technical report highlighting DSA architecture and asynchronous RL. The OpenClaw multi-agent framework went viral but triggered Anthropic account bans over OAuth usage. Infrastructure items include CoDA-GQA-L compressing KV cache from 160GB to 136MB, GPU MODE kernel optimization results on B200, and confirmation that FP4 formats only shine on Blackwell GPUs. Research highlights cover the Moltbook 2.6M-agent social simulation and the Every Eval Ever benchmark standardization effort. Product news includes Sonnet 4.6 rollout across Cursor, Perplexity, and Cline, PolyAI's $200M raise, and Mistral's acquisition of Koyeb.

Easy AI Daily | February 18, 2026

A translated digest of AI industry news originally published on zhichai.net.

Models and Capabilities

Anthropic launches Claude Sonnet 4.6: 1M context, near-Opus quality, new cost trade-off

Anthropic released Sonnet 4.6, calling it 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 scored ELO 1633 (first place) on third-party GDPval-AA, but consumed roughly 4.8x the tokens of Sonnet 4.5—more than Opus 4.6—so total cost on complex tasks may be higher. Published 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.

Links: Official release | GDPval-AA analysis | Reddit discussion

Qwen3.5-397B-A17B: Alibaba open-sources a 397B MoE with multimodality and 1M context

Alibaba released Qwen3.5-397B-A17B: 397B total parameters with 17B active (MoE), native 262K context expandable to 1M, supporting 200+ languages plus image, text, and video. Alibaba claims performance near GPT-5.2, Claude Opus 4.5, and Gemini 3 Pro on instruction following, multilingual knowledge, and video understanding. GGUF quantizations exist for local runs on high-memory Macs or multi-GPU setups, though VRAM demands are extreme.

Links: Official blog | Hugging Face | GGUF

GLM-5 technical report: DSA, async 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 DSA architecture, asynchronous RL infrastructure, and RL algorithms for agent scenarios. Some developers report underperformance in certain IDE/plugins, likely due to integration and inference stack issues.

Links: arXiv:2602.15763

Tiny Aya: Cohere's 3.35B multilingual small model for phones

Cohere Labs released Tiny Aya, a 3.35B-parameter, 70+ language open model targeting phones and edge devices, reportedly trained on just 64 GPUs. The goal is practical deployment rather than top benchmark scores.

Agents and Tooling

OpenClaw virality and backlash: The multi-agent computer-control framework spread rapidly, but users connecting via Claude OAuth reported account bans, as Anthropic considers third-party desktop OAuth access a TOS violation. Security concerns remain around giving LLMs local read/write, browser, and credential access.
Harness engineering: LangChain advocates turning production traces into automated eval loops (LangSmith Insights, TerminalBench) rather than relying on offline benchmarks.
MCP payments proposal (SEP-2007): A proposal to let MCP servers signal paid tools (starting with X402); maintainers are cautious and prefer URL-based approaches for now. See PR #2007.
DirectShell: Uses OS accessibility APIs as a universal app interface instead of screenshots and coordinate clicks, enabling scriptable control of any desktop app. GitHub

Infrastructure and Hardware

GPU MODE kernel results: ~350→368 TFLOPS from custom persistent CUDA matmul kernels; 5.7x speedup from FlashInfer's MoE baseline on B200; discussions of FP8/FP4 numerical error and profiling tools.
CoDA-GQA-L: Bounded-memory attention fixing KV cache at 384 slots per layer (256 recent tokens, 64 landmark tokens, 64 EMA summaries), compressing KV from 160GB to 136MB. Paper
MXFP4/NVFP4: FP4 gains require Blackwell (compute capability ≥12.0) native Tensor Cores; on Ampere, simulated FP4 can be slower than FP8.
Cloud outages: OpenRouter mass 401s, Perplexity API auth failures with active balances, and Hugging Face endpoints returning 500s despite green status pages—teams should keep endpoint-rebuild contingency scripts ready.

Research

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 compression with restorable references, reportedly more stable than Claude Code from 32K to 1M context on OOLONG.
Moltbook: A simulation with 2.6M LLM agents producing 300k posts and 1.8M comments found macro-level culture converges and stabilizes, but individual agent influence is statistically negligible noise.
Every Eval Ever: The EvalEval consortium proposes a BIDS-like standard for LLM benchmark metadata, motivated by issues like Cybench's non-randomized flags inflating scores.
Preventative steering: Anthropic's persona vectors inspire adding interference in representation space to test whether models hold targets—and using that as data augmentation.

Products and Industry

Sonnet 4.6 rollout: Live in Cursor, Windsurf, Cline (free trial; 70% of developers preferred it in internal A/B), Perplexity Pro/Max, and OpenRouter.
PolyAI: Raised $200M (Nvidia, Khosla participating) and launched Agent Studio Lite—paste a URL, get a voice agent in 5 minutes.
SpeakType: A fully local, open-source Whisper-based dictation app for Mac.
Manus Agents: A personal agent with long-term memory and Gmail/Notion integration, accessed via Telegram.
Food Truck Bench: 12 LLMs each given $2,000 and a food truck for 30 days—only 4 survived; Claude Opus 4.6 earned ~$49k, GPT-5.2 ~$28k. Models that took loans almost always went bankrupt.
Mistral acquires Koyeb: Moving from model vendor to full model-plus-compute cloud provider.
Grok 4.20: 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 political stances.
Anthropic and the military: Claude cleared for some US military use with two red lines: no mass surveillance, no autonomous weapons.
Waymo: 6th-gen Robotaxi platform costs ~$70k per vehicle (expected to halve by 2028), with 500k+ weekly driverless rides growing 3x annually.

Policy and Security

OpenClaw OAuth bans highlight that desktop agents shouldn't piggyback on web OAuth—use official APIs or self-hosted models.
Dark web listings: GitLab projects with maintainer access auctioned ($200–$2,000); "unlimited AI" services on Tor raise malware and abuse concerns.
API key leakage: An OpenRouter user's key leaked despite .gitignore and was drained of $10 in 20 minutes—practice least privilege, rotation, and repo-wide secret scans.
Cybench flaw: Fixed flags from known CTFs let models recall answers; randomizing flags sharply dropped success rates, showing how benchmark design flaws inflate security evals.

---

📌 Source: Easy AI Daily (zhichai.net)