Grok Build CLI $300/mo launched in early beta (70.8% SWE-Bench); Cursor reaches $3B ARR, SpaceX acquisition, xAI partnership; RecursiveMAS 2.4x. MCP enables 80-100 tool-call agents. Microsoft Webwright: 60.8% Odysseys/86.7% Online-Mind2Web on GPT-5.4. BeSafe-Bench: no agent passes 40% safe completion. Starlette vulnerability (BadHost) exposes millions of AI agents. Exabase M-1 achieves 96.4% on LongMemEval at 4-6x cheaper. AC/DC framework emphasizes verification bottleneck. AlphaProof Nexus solves 9 Erdős problems with Lean-verified proofs. AI security shift: researchers argue for system-level controls; ADR framework and benchmark introduced. Prompt/retrieval/eval debt reshaping enterprise AI risk. First large-scale study of formal theorem proving by LLM agents noted. Langfuse evals pipelines. Claw-Anything benchmark (GPT-5.5 34.5% pass@1), LARA EU compliance benchmark (Claude Opus 4.7 ~54%), POSTTRAINBENCH for autonomous post-training agents, Cisco multi-turn attack study (8-88% ASR), MATCHA paper on eval contradictions, noisy LLM evaluators still effective, EAGLE 3.1 speculative decoding (2.03x throughput on Kimi K2.6), 42Crunch API security for Claude Code, Mirage mounts cloud services as local folders. AutoTTS automates test-time scaling strategy design, cutting token usage 69.5% while maintaining accuracy. Mistral launches Vibe agent for work and code. OWASP Top 10 LLM risks updated. FAX framework for faithful agentic XAI with new benchmark. Cost optimization guide (AI.cc) shows 80% savings via tiered routing, prompt compression. Dynamic workflows demonstrated with 50+ parallel agents for due diligence. Agent observability deep-dive: stack traces, metrics, 15-min debug loop. Google's 1,000-agent marathon demo with ADK/A2A/Redis pub-sub. Token discipline shift: tokenmaxxing dying, cost optimization now key. Graphon intelligence layer pre-processes data relationships outside LLM, potentially reducing context costs. DMind benchmark for Web3 safety reveals critical gaps in GPT-5, Claude, Gemini on smart contract auditing and tokenomics; KDD 2026. Prompt injection via hidden HTML elements in Markdown links turns ChatGPT into attack channel. FreeLLMAPI tool aggregates 14 free tiers into one endpoint, offering up to 1B free tokens/month. OpenRouter hits $1.3B unicorn with Alphabet leading, processing 25T tokens/week (5x in 6 months). Code quality study: GPT-5.4 generates 1.2M lines for 4,444 assignments (vs GPT-4o's 250k), raising verbosity/maintainability concerns; Claude Sonnet 4.6 has 300 security issues per million lines; ACDC framework for verification bottleneck. Tether AI open-sources TurboQuant, reducing KV cache memory by 5x. Headroom context optimization layer claims 60-95% token reduction. Amazon shuts down internal AI leaderboard after employees cheated. LiveBrowseComp reveals AI search agents guess before searching, scores drop 25-40 points. Grok Build 0.1 API launched with parallel architecture and MCP support. Microsoft MAI-Thinking-1 reasoning model announced, enterprise-targeted. WBench benchmark for interactive world models shows navigation ability independent of other skills. X-Stream benchmark tests MLLMs on multi-stream video understanding; SOTA models only ~50%, exposing major gap. NVIDIA RTX Spark with 128GB unified memory enables local 120B agent deployment, shifting edge AI. ESPO (Early-Stopping PPO) cuts wasted tokens by >20% on math benchmarks, aligning with token discipline. Google Managed Agents in Gemini API simplifies agent creation with single API call. Cost concerns: Microsoft data shows AI more expensive than humans; mystery firm burns $500M on Claude in one month. PEFT scaling paper reframes adapters as persistent local state for personal models, aligning with token discipline. GenAI trust deficit: only 31% of AI use cases reach production, 1/4 achieve expected ROI; PocketOS database deletion incident. NITP training technique achieves 5.7% MMLU-Pro gains on 9B MoE with zero inference cost. RoboSemanticBench exposes VLA models' semantic grounding failures. Copilot metered billing backlash: users report 2x cost increases and inferior tooling vs Cursor/Claude Code, signaling market shift. Harness variability (7-32 point benchmark gaps) confirmed as critical factor in agent evaluation. Microsoft launches MAI family at Build 2026: MAI-Thinking-1 reasoning model, MAI-Code-1-Flash (137B, 51% SWE-bench pro), MAI voice/image models. Also launches Scout, an OpenClaw-inspired personal assistant for M365 with policy conformance. GitHub Copilot app goes agent-native with My Work, canvases, sandboxes. Microsoft ASSERT framework generates AI behavior tests from text. LongAttnComp context compression matches full accuracy on code debugging, transfers across families. Geometric Latent Reasoning (GLR) reduces token usage by replacing early explicit reasoning with continuous latent steps, tested on Qwen3 math. NVIDIA releases Cosmos 3, a two-tower MoT foundation model for physical AI, open-source with Nano (16B) and Super (64B) scales, leading VANTAGE-Bench, TAR, Physics-IQ, PAI-Bench. $500M AI bill shock from unchecked Claude usage underscores tokenmaxxing death and cost optimization imperative. Big-T Notation framework from Adobe engineers offers practical token efficiency lens. Mollick critique notes AI 'everything apps' still chat+IDE hybrids, missing non-linear knowledge work. New research: Scaling the harness (system scaffolding) identified as next major bottleneck in agentic AI; deception probe fragility revealed (probes collapse under style shift); multi-agent scaling behavior studied (adding more agents may not always help). Microsoft ASSERT framework turns AI-agent policies into executable tests with OpenTelemetry traceability, 80-90% judge-human agreement, MIT license, cross-framework support. Headroom context compression claims 95% reduction, directly addressing tokenmaxxing death. World Models Meet Language Models paper (PF-OPSD) shows 10.6% and 10.9% gains on new benchmarks, promising for grounding LLMs in physical prediction. Self-improving LLMs with bidirectional evolutionary search framework shows consistent accuracy gains on logical reasoning benchmarks with Gemma and Yama models. Uber caps AI spend at $1,500/mo, signaling enterprise pricing tolerance and tokenmaxxing death. Hyper (YC P26) launches 'company brain' for agentic development, addressing agent context bottleneck. Microsoft MAI-Thinking-1 (35B active params, 256K context, low-token cost) enters reasoning model race with transparency emphasis. Microsoft Build 2026 also introduces Aion on-device models (1.0 Instruct and 1.0 Plan) for local AI. Weaviate launches Engram, a managed memory service for agents that actively maintains memory. Big Techday talk by Florian Brand (Prime Intellect) critiques LLM benchmarks for agentic systems, highlighting implementation differences and infrastructure challenges. OpenAI proposes mandatory third-party evaluations for advanced AI models, diverging from White House voluntary framework. BenchEvolver introduces solution-centric evolution for benchmark tasks; LiveCodeBench-Plus restores 27.5-62.6% Pass@1 range for frontier coding models. MemTrain self-supervised context memory training achieves 17.67-point gain on long-text QA via GRPO. TELBench and DRIFT provide span-level error localization for deep-research agents, enabling process-level debugging. Coralogix raises $200M for AI agent observability, signaling infrastructure maturity. Perplexity's Search as Code approach for agents improves token efficiency. Cosmos 3 paper released with architecture details and benchmark leadership. Tool-calling benchmark sensitivity paper reveals leaderboard unreliability due to implementation choices; RL training speedup 2.6x by skipping zero-variance prompts. On-policy distillation highlighted as impactful at frontier; teacher quality gap matters more than method. Alphabet raises $85B, guiding $180-190B capex in 2026, 300x token volume growth to 3.2 quadrillion/month; serving costs down 78% in 2025, 30% more since Gemini 3. Coralogix Series E $115M (not $200M) for AI agent observability. Practical eval framework from Xelix for production agents. LangSmith on AWS evaluation patterns for deep agents. Neo Research launches as Asia's first independent frontier AI safety evaluation lab. OpenAI publishes a methodological playbook for third-party evals, addressing harness variability and reward hacking. NVIDIA Nemotron 3 Ultra (550B MoE, 55B active) released as open model optimized for agent orchestration, leading PinchBench/IFBench/Ruler @1M but trailing on Terminal-Bench. Microsoft Agent Control Specification proposes policy-as-code for agent governance with SDK plugins. VaSE (Value-Aware Stochastic KV Cache Eviction) targets reasoning model CoT bloat. NVIDIA/Microsoft sparsity research could bring million-token context to consumer GPUs. Satya Nadella podcast emphasizes private evals as critical IP and multi-model harnesses. SynthTraces tool for generating synthetic coding agent traces released. Andrew Ng course on efficient LLM serving. ZeroDrift raises $10M for AI compliance layer. Airbnb launches in-house AI lab. Latest additions: AdaPlanBench reveals frontier models only 67.75% on adaptive planning under user constraints. MLEvolve achieves SOTA on MLE-Bench with half runtime via Progressive MCGS and Retrospective Memory. Meta-Cognitive Memory Policy Optimization achieves 97.1% performance at 1.75M tokens using belief entropy. EvoDS improves data science agents 28.9% over open-source SOTA. Shadow Price paper introduces CLEAR for optimal budget allocation, up to 3x accuracy improvement under budget constraints. WLA-0 achieves SOTA on RoboTwin2.0 with 2B params, 40ms inference. Continual Experience Internalization paper provides design principles for self-evolving agents. ArcANE benchmark for role-playing agents shows Arc condition outperforms. Microsoft Build 2026 financial analysis highlights Maia 200 chip and MAI models. NVIDIA Cosmos 3 open-source physical AI model released with architecture details.
