Meta introduced Muse Wonder, a frontier agent model pre-trained with native browser operation capabilities and state-of-the-art performance across autonomous benchmarks.
This brief is compiled automatically by AI and may contain errors. Check the linked sources.
Model Releases & Updates
Muse Wonder & Muse Wonder FullMMeta
Bottom line：Meta introduced Muse Wonder, a frontier agent model pre-trained with native browser operation capabilities and state-of-the-art performance across autonomous benchmarks.
Benchmark Performance：Debuted at #12 on the real-world Agent Arena leaderboard with a +6.4% net improvement, demonstrating frontier-grade capability on multi-step personal and professional tasks.
Native Web Navigation：Trained directly on browser interactions, allowing the model to complete end-to-end shopping and workflows on standard websites without custom integrations.
Bottom line：xAI unveiled Grok Voice 3.0, a speech-to-speech foundation model offering sub-second latency and automated call resolution for enterprise voice workflows.
Benchmark Leadership：Captured the top rank on the Artificial Analysis Speech-to-Speech Index with a score of 86.9, outpacing GPT-Live-1 at 83.2 and Gemini 3.8 Live at 82.6.
Latency and Economics：Delivers an average first-audio latency of 0.81 seconds with a 94.8% task success rate, priced at $0.08 per minute via the Realtime API.
Enterprise Deployment：Currently deployed in production at SpaceX, autonomously resolving approximately 70% of inbound Starlink customer support calls.
Bottom line：HeyGen launched HeyGen Voice, a video-first speech model engineered to align pacing, expression, and vocal cadence with scene-level acoustics.
Blind User Preference：Ranked #1 on the Artificial Analysis TTS leaderboard after 42,875 blind evaluations across five languages, winning 77.9% of head-to-head comparisons against Cartesia.
Pricing & Modes：Released in standard and low-latency Turbo versions on OpenRouter and direct API, starting at an introductory rate of $7.50 per million characters.
Bottom line：Fish Audio published open weights for Drama 3, a 35-billion parameter Mixture-of-Experts speech synthesis model.
Architecture：Utilizes a 35B parameter Mixture-of-Experts architecture that activates roughly 3 billion parameters per token for compute-efficient real-time generation.
Open Availability：Full model weights have been released openly on Hugging Face alongside enterprise inference endpoints on the Fish Audio API.
Bottom line：NVIDIA introduced Nemotron 3.5 Lightning, an open-weight hybrid Mamba-MoE model designed for low-latency reasoning and long-running agent workflows.
Mamba-MoE Hybrid Backbone：Combines 30B total parameters with 3B active parameters per token by interleaving Mamba-2 state-space layers, MoE blocks, and targeted attention layers.
Context & Availability：Provides a 1-million-token context window with native tool calling, hosted for free experimentation via Vercel AI Gateway.
Bottom line：OpenAI rolled out ChatGPT Space, a shared workspace for teams to co-create living documents and orchestrate agent tasks with synchronized external tools.
Collaborative Canvas：Allows teams to transform conversational threads into persistent documents with real-time human and agent editing, inline comments, and task tracking.
Automated Synchronization：Enables users to instruct ChatGPT to continuously update workspace pages with inbound data from connected developer and business tools.
Bottom line：Anthropic shipped automatic idle compaction in Claude Code to compress dormant sessions before prompt caching expires, curtailing re-reading token costs.
Proactive Compaction：Automatically summarizes sessions exceeding 200,000 tokens after approximately 54 minutes of inactivity before the 1-hour prompt cache expires.
Configurable Controls：Developers can disable the feature or adjust the minimum token threshold using the CLAUDE_CODE_IDLE_COMPACT_MIN_TOKENS environment setting.
Bottom line：JetBrains released a preview of cloud agent sandboxing in Air, providing isolated remote environments for autonomous coding agents.
Remote Isolation：Runs AI agent tool use, file reads, and code execution within JetBrains-managed cloud environments rather than touching local developer machines.
Multi-Surface Supervision：Enables developers to initiate coding tasks in IntelliJ IDEA and monitor execution state across browser dashboards and mobile devices.
Bottom line：Manus introduced Teams and Enterprise tiers, bringing shared agent templates, unified billing, and access controls to its autonomous agent platform.
Shared Workspaces：Enables cross-functional teams to share contextual workflows, execution histories, and custom agent playbooks across marketing, product, and engineering.
Enterprise Governance：Integrates centralized user provisioning, administrative audit logs, and domain-level data governance policies.
Bottom line：Databricks added native SQL predictive AI functions to automate anomaly detection, causal analysis, and time-series forecasting inside the data warehouse.
SQL-Native Functions：Introduces ai_top_drivers, ai_detect_anomalies, ai_predict_value, and ai_predict_class directly into standard SQL queries without bespoke ML pipelines.
Forecasting Improvements：Upgraded ai_forecast with enhanced accuracy and lower memory consumption across large-scale transactional datasets.
Modal Reaches $1B Run Rate and Raises $750M Series DMModal
Bottom line：Serverless AI compute provider Modal closed a $750 million Series D at a $15.75 billion valuation after growing annualized revenue 15x to exceed $1 billion.
Financial Milestone：The financing round was led by Accel with participation from Kleiner Perkins, valuing the startup at $15.75 billion.
Infrastructure Specialization：Attributes growth to its proprietary container scheduler and virtualized file system designed to deliver sub-second cold starts for AI agents and GPU workloads.
Eight US States Sue OpenAI Over Minor Safety and Consumer DeceptionSState Attorneys General
Bottom line：Attorneys general from eight US states filed lawsuits against OpenAI alleging deceptive trade practices and inadequate safety guardrails for minors using ChatGPT.
Multi-State Action：Lawsuits were lodged independently across Iowa, Kansas, Nebraska, and five other states under local consumer protection statutes.
Core Allegations：The complaints allege OpenAI failed to prevent underage access, collected minor user data without parental consent, and downplayed systemic risks.
Mistral AI Acquires Design Studio New MaterialMMistral AI
Bottom line：Mistral AI acquired European digital product and AI interface design agency New Material to strengthen its frontend applications and enterprise deployment stack.
Team Absorption：New Material’s design and engineering group joins Mistral directly to lead user interface and product experience development.
Domain Portfolio：New Material previously delivered custom AI systems for clients including Ford, Rafael Nadal, the Wellcome Trust, and AmLaw 50 law firms.
40 State Attorneys General Intervene in Character.AI BankruptcyOOffice of the New York AG
Bottom line：A bipartisan coalition of 40 state attorneys general intervened in Character.AI’s bankruptcy case to block the unvetted sale or transfer of user chat history.
Bipartisan Filing：Led by New York AG Letitia James, 40 state officials submitted a formal notice of appearance in Delaware bankruptcy court.
Privacy Mandate：Demands strict enforcement of data protection terms to prevent user dialogue logs and minors’ personal identifiers from being sold to third parties like Disney.
AutoEnvScaling: Synthetic RL Environments Designed by AgentsMMicrosoft Research
Bottom line：Researchers developed AutoEnvScaling, an automated framework where a proposer agent designs terminal-based RL environments directly from model rollout logs.
Benchmark Gains：Lifted Qwen3.6-35B-A3B performance on Terminal-Bench 2.1 from 41.1 to 53.3 through iterative training on self-generated environments.
Synthesis Efficiency：Building environments inside an interactive terminal yielded up to 25x more valid environments at a 75% lower cost per instance compared to prompt-only generation.
Hybrid Reasoning Agents: Latent Thinking with Text ActionSStanford University & Together AI
Bottom line：A new paper introduces Hybrid Reasoning Agents, which reason internally across latent representations and only convert to text when executing tool actions.
Efficiency & Speed：Reduces token consumption by 35.5% to 54.7% and increases execution speed by 1.8x to 2.6x relative to standard text-based reasoning agents.
Evaluation Results：Surpassed frontier LLM agents by an average of 10.4% across 11 benchmark suites covering science, medicine, and agentic workflows.
Deep Review: Multi-Agent Technical Document AnalysisGGoogle Research
Bottom line：Google Research detailed Deep Review, an automated multi-agent framework designed to analyze claims, audit code, and review complex scientific and financial documents.
Cross-Domain Validation：Expanded from reviewing computer science manuscripts to auditing legal contracts, regulatory filings, and financial balance sheets.
Rigorous Fact Checking：Employs specialized sub-agents to trace empirical claims back to source tables, test reproduced algorithms, and flag unverifiable deductions.
Bottom line：An MIT research team published a machine learning post-processing model in Nature that doubles the predictive accuracy of weather forecasts two to six weeks out.
Systemic Correction：Learns historical error patterns in legacy physics-based simulations, beating climatological baselines consistently across multi-week horizons for the first time.
Real-World Adoption：The open-source, low-compute layer has been adopted by the Kenya Meteorological Department to improve early detection of monsoon-related flooding.
$1M Vercel Sandbox Challenge Discovers Linux Kernel Zero-DayVVercel
Bottom line：Vercel disclosed the findings of its $1M security challenge, revealing that a researcher used an AI swarm to uncover an upstream Linux kernel vulnerability that breached microVM boundaries.
Zero-Day Impact：Paulos Yibelo discovered that crafted networking packets from within a standard microVM could trigger host crashes, memory leaks, and host code execution.
Program Scope：Paid out $242,000 across 1,285 submissions, confirming that no customer data was compromised and that upstream Linux patches have been coordinated.
AA-Robotics v0.1 Physical Arm BenchmarkAArtificial Analysis
Bottom line：Artificial Analysis launched AA-Robotics v0.1, a benchmark evaluating how effectively frontier reasoning models command a 6-axis physical robot arm across 8 tabletop tasks.
Real-World Execution：Tests frontier models inside an agentic coding harness where they plan, generate control commands, and manipulate real desktop objects.
Early Leaderboard：GPT-6 Astra captured early top rankings in spatial trajectory planning and automated visual error correction.
