Easy AI Daily News | 2026-02-18

A digest of AI industry news for February 18, 2026, originally published on zhichai.net.

Models & Capabilities

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

Anthropic released Sonnet 4.6, billed as its strongest Sonnet model, with upgrades across coding, computer use, long-context reasoning, and agent planning. It offers 1M token context (beta) at the same pricing as Sonnet 4.5. Third-party GDPval-AA evaluation ranked it first with ELO 1633, but achieving that required ~4.8x the token consumption of Sonnet 4.5 (also above Opus 4.6), so total cost on complex tasks may be higher. Reported scores: 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 outputs were quickly fixed.

Links: Official release | Anthropic tweet | Benchmarks roundup | Artificial Analysis on GDPval-AA and token usage

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

Alibaba released Qwen3.5-397B-A17B: a MoE with 397B total and 17B activated parameters, native 262K context extendable to 1M, 200+ languages, and image/video multimodality. Alibaba claims near-GPT-5.2, Claude Opus 4.5, and Gemini 3 Pro performance on instruction following, multilingual knowledge, and video understanding benchmarks. GGUF quantized versions exist for high-memory Macs or multi-GPU local setups, though VRAM demands make it best suited to cloud/cluster deployment.

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 performance on benchmarks like WeirdML, approaching or exceeding some closed models. The report highlights the DSA architecture, asynchronous RL infrastructure, and RL algorithms for agent scenarios. Some developers report weaker-than-claimed results inside certain IDEs/plugins, possibly due to integration and inference stack issues.

Links: arXiv:2602.15763 | Demo video

Tiny Aya: Cohere's 3.35B multilingual small model, "runs on phones"

Cohere Labs released Tiny Aya, a 3.35B-parameter, 70+ language open model targeting mobile/edge deployment. Reportedly trained on just 64 GPUs with a detailed technical report, aiming at practically deployable multilingual assistants rather than top benchmark scores.

Links: Announcement | Model card

Agents & Tooling

OpenClaw goes viral — and hits trouble: bans and security concerns

OpenClaw, a multi-agent framework controlling real computers, spread rapidly for project management, CRM, and browser automation. But users connecting via OAuth to Claude reported account bans, with TOS interpretations treating third-party desktop control tools as "reverse-engineered/unauthorized" access. Security concerns center on granting LLMs read/write access to local machines, browsers, and credentials, where prompt injection or malicious skills could cause serious damage.

Links: OpenClaw showcase | ClawRecipes

LangChain pushes "Harness Engineering": turning real logs into agent test arenas

LangChain emphasizes using production traces as automated evaluation and self-verification loops for iterating agent workflows, plus TerminalBench-style terminal tasks and scheduled LangSmith Insights evaluations.

Links: LangSmith Insights | OpenAI blog post

MCP discusses paid tools: adding payment signals to the protocol

MCP proposal SEP-2007 would let MCP servers signal "payment required" (starting with X402) so agents can call paid tools within budgets. Maintainers are cautious, preferring URL-redirect approaches initially; not a near-term core priority.

Links: PR #2007

DirectShell: turning the OS accessibility layer into a "universal app interface"

DirectShell controls apps via system accessibility APIs instead of screenshots and coordinate clicks, abstracting the accessibility layer into a unified DOM/control tree so any desktop app becomes scriptable — potentially a key technology for PC agents.

Links: Design blog | GitHub

Infrastructure & Hardware

GPU MODE: squeezing TFLOPS from B200 / RTX cards and MoE inference

Engineers shared hand-written matmul and FlashInfer tuning details: ~350→368 TFLOPS on custom CUDA persistent kernels, 5.7x speedup with FlashInfer's MoE baseline on B200, plus discussions of FP8/FP4 numerical error, Nsight Compute vs CUDA events usage, and approaching peak achievable FLOPs.

Links: matmul_V3 source | FlashInfer baseline | Achievable FLOPs reference

CoDA-GQA-L: KV cache compressed from 160GB to 136MB via "bounded memory attention"

Each layer's KV cache is fixed at 384 slots: 256 recent tokens, 64 landmark tokens, 64 exponentially-moving-average summaries. A practical path for 1M-context apps (search, long-form writing, long-task agents) that can't afford the memory bill.

Links: Paper (Zenodo) | Code

MXFP4 / NVFP4: FP4 only shines on Blackwell, older cards lose

Unsloth and Modular confirm FP4 formats like MXFP4 need Blackwell (compute capability ≥12.0) native Tensor Cores; on Ampere-era cards emulation is often slower than FP8. NVFP4 is Modular MAX's current optimization focus.

Links: Unsloth notes | Modular forum

Cloud inference outages: OpenRouter/HF/Perplexity 401s and 500s

OpenRouter suffered widespread 401s (war-room fix), Perplexity API users hit persistent 401s despite credit, and HF Inference Endpoints returned 500 "service panicked" even with green status pages, forcing endpoint rebuilds. Practical lesson: prepare scripts to fully rebuild endpoints when they fail inexplicably.

Links: OpenRouter status | HF status

Research & Methods

Agent World Model / Lossless Context Management

Two research directions: Agent World Model (1,000 executable environments, 35,062 tools, 10K tasks for RL training of tool-use agents) and Lossless Context Management (LCM/Volt), using hierarchical DAGs with resumable references — reportedly more stable than Claude Code from 32K to 1M context on OOLONG. Common theme: structured memory around tools rather than brute-force long context.

Links: Agent World Model | LCM/Volt on OOLONG

Moltbook: a 2.6M-agent social experiment — macro culture stabilizes, individual influence is noise

In a multi-agent social simulation with 2.6M LLM agents producing 300K posts and 1.8M comments, overall culture and discourse style converged and stabilized, but individual agents' "influencer" effects were statistically negligible — a reality check for "stack more agents = AGI" narratives.

Links: Discussion

Every Eval Ever: a BIDS-style standard for LLM evaluations

The EvalEval consortium proposes unified data/metadata standards for LLM benchmarks. Motivation: benchmarks like Cybench showed design flaws (non-randomized flags inflating scores) that can drastically reorder model rankings once fixed.

Links: Every Eval Ever | Cybench

Preventative steering: adding interference in representation space

Discussions around Anthropic's persona vectors: apply a steering vector to intermediate representations and test whether the model still reaches its goal — potentially a form of representation-space data augmentation.

Links: arXiv:2602.15029

Products & Applications

Sonnet 4.6 rollout: Cursor, Perplexity, Windsurf, Cline all onboard within a day

Cursor finds it clearly better than 4.5 on long tasks though slightly below Opus 4.6 in "intelligence"; Cline reports 70% of developers preferred 4.6 in internal A/B tests (with short-term free access); Perplexity integrated it into Deep Research/browser agents. Trend: Sonnet 4.6 as the cost-effective workhorse, Opus for maximum quality.

Links: Cursor | Windsurf | Perplexity

PolyAI raises $200M, launches 5-minute voice agent builder Studio Lite

Voice customer-service company 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.

Links: Announcement

SpeakType: fully local open-source dictation for Mac

A Whisper-based, fully offline Mac dictation app, open source, focused on privacy. Community questions center on RAM/CPU usage, VAD preprocessing, and comparisons to tools like Handy.

Links: GitHub | Product page

Manus Agents: a long-memory personal agent inside Telegram

Manus Agents offer a personal agent with long-term memory and tool integration (Gmail, Notion) via Telegram, generating videos, slides, and web pages.

Links: Announcement

AI runs food trucks: 12 models, $2,000 each, only 4 survived

A 30-day simulation gave 12 LLMs $2,000 and a food truck; only 4 survived. Claude Opus 4.6 earned ~$49K, GPT-5.2 ~$28K; nearly every model that took loans went bankrupt — echoing Vending-Bench 2's lesson that aggressively expanding/borrowing agents die first in reality.

Links: Experiment thread | Vending-Bench 2

Industry & Company News

Mistral acquires Koyeb: building its own inference cloud

Mistral plans to acquire serverless platform Koyeb, folding its platform and team into Mistral Compute — moving from "API-only model company" toward a model-plus-compute cloud vendor like OpenAI, Anthropic, and xAI.

Links: Announcement

Grok 4.20: mocked as "four Grok 4.1 agents in a trench coat," criticized for political bias

xAI released Grok 4.20 claiming a new architecture, but log inspection shows the model ID is still grok-4-1-thinking-1129. Reddit users also report it echoing Musk's positions on gender/political topics and underperforming peers.

Links: Discussion

Anthropic allows military use of Claude with two red lines

Anthropic agreed to limited US military use of Claude, explicitly prohibiting mass surveillance and autonomous weapons.

PolyAI and Waymo: AI absorbing customer service and robotaxis

PolyAI serves clients like Marriott, raised $200M, and is building a self-serve voice-agent platform. Waymo's sixth-gen robotaxi platform costs ~$70K per vehicle (expected to halve by 2028), with over 500K driverless weekly rides growing ~3x annually.

Links: Waymo economics

Policy, Governance & Safety

OpenClaw + OAuth bans: third-party desktop agents treated as "reverse engineering"

Anthropic bans on OpenClaw users connecting via Claude OAuth reflect TOS clauses on unauthorized third-party access. Lesson for desktop-agent teams: don't rely on vendor web OAuth as a free backend; use official APIs or self-hosted models.

GitLab project auctioned on the dark web, "unlimited AI" on Tor

Dark web actors auctioned three GitLab projects with maintainer access (PHP/Laravel stack, starting bid $200, flash sale $2,000); an "unlimited AI" service appeared on Tor, with warnings about malware and Claude abuse. LLM tech and gray/black markets are increasingly overlapping.

Links: X post

API key leaks again: OpenRouter user drained $10 in 20 minutes

A leaked OpenRouter key (despite being in a gitignore file) was drained of $10 in 20 minutes via a third-party "Cloud Code" environment. Standard advice applies: least privilege, regular rotation, repo-wide secret scans before launch.

Cybench's non-randomized flags: one design oversight inflates defense evals

Early Cybench versions used fixed flags from well-known CTFs, letting models recall them from training data and inflate success rates; randomizing flags caused scores to drop sharply. Rigorous scenario design is essential for security/red-team benchmarks.

Links: Cybench

---

📌 Source: Easy AI Daily