Cover Story · Model Release
Google Launches Gemini 4 Argon Frontier Model
A new frontier model tuned for coding, enterprise knowledge and complex workflows — Google's sharpest bid yet to rejoin the top tier of AI.
Google DeepMind has launched Gemini 4 Argon, a new frontier model aimed squarely at the workloads that define modern enterprises: coding, structured enterprise knowledge and long, multi-step workflows. The release is being read across the industry as Google's clearest attempt in years to reclaim a seat at the very top table of frontier models. Early reactions from the research community point to strong performance on reasoning-heavy benchmarks, with particular attention on how Argon handles long-horizon agentic tasks where earlier Gemini generations had lagged rivals. Coming on the heels of a rapid cadence of model announcements, the launch resets expectations for the fall model season and puts direct pressure on OpenAI's GPT-6 Astra and Anthropic's Claude line.
Claude Code Opens Mod Customization
Users can now modify Claude Code's behavior and interface with a few lines of TypeScript, swap in their own features, and distribute Mods as plugins installed via /plugin in the CLI or desktop app. The change turns the coding agent into an extensible platform, letting teams encode their own conventions, tools and guardrails directly into the editor experience. Anthropic's move follows a broader industry shift toward agent customization, and arrives as rival coding tools race to open their own plugin and extension surfaces.
DeepMind Launches SynthID Bio Biological Watermark
A new family of watermarking methods made for AI-generated biological designs can embed an imperceptible signature directly into protein sequences without affecting biological function — a world first.
GPT-6 Astra Ultrafast Gets 8x Speedup on NVIDIA
NVIDIA says GPT-6 Astra Ultrafast runs on Blackwell GPUs, and with ongoing inference optimizations for OpenAI models, delivers up to 8x faster responses across code generation, tool calls and interactive applications.
Perplexity Open-Sources 27B Decision Model and Decisions API
Perplexity has open-sourced pplx-decider-27b, a state-of-the-art multimodal decision model, and is offering it through a new Decisions API at 4 cents per million input tokens with free output — promising further price cuts in the coming days.
Runway Unveils Real-Time Video OS Continuum
Runway Labs has shown an early preview of Project Continuum, an operating-system research application built around real-time video interfaces, demonstrating four new modes of interaction including Portals, visual thinking and responsive video interfaces.
Black Forest Labs Releases FLUX 3 Image
Built on the multimodal FLUX 3 backbone, it excels at multi-turn editing, generates natively at up to 4K, and supports bounding boxes for precise localized editing and generation; an open-weight variant is coming.
Tavus Releases Griffin, Model Passing Video Turing Test
Tavus says Griffin is the first model to pass the video Turing test, with 48% of people who spoke to it live believing they were talking to a human.
Looped Diffusion Transformer: Small Model Beats 6.5x Larger
The paper repeats a shared Transformer block within each denoising step to scale compute without adding parameters. With deep supervision and self-modulating attention to stabilize the loop, a 260M model outperforms models 6.5x larger on text-to-image benchmarks while using 4.9x less inference compute.
GLM 5.3 Series Launches on Cursor
GLM 5.3 and GLM 5.3 Flash are now available on Cursor; the company says GLM 5.3 Max is the highest-scoring open-weight model on CursorBench 4.0.
The critical distinction between base LLMs and modern reasoning models is not symbolic tool use — it is the switch from a transductive paradigm to an inductive paradigm.
— François Chollet
Cloudflare Open-Sources First In-House Models clef
The Workers AI team released clef and clef-flash on Hugging Face under Apache 2.0, positioned as an alternative to Jev.
9TB First-Person Video Dataset Open-Sourced
Before winding down, eidon.ai open-sourced 9TB of first-person video for robot learning and embodied intelligence research.
LlamaIndex Launches Document Extraction Agent Extract v2.5
A frontier agent series for document extraction focused on cost-controlled, high-quality structured output.
Synthesia Launches Interactive AI Avatars Sessions
A set of interactive AI avatars for real-time coaching, interviews and guided conversations.
Alibaba Wan3.0 Tops Video Generation Leaderboard
Ranked first overall on the new Artificial Analysis leaderboard as the generation quality bar keeps rising.
Altman: 6.1 Sol Is Fastest-Growing Model, Load Improved
Sam Altman says 6.1 Sol is OpenAI's fastest-growing model ever, and slow responses under load have now improved.
Gemini 4 Argon Matches GPT-6 Astra at 60% Cost
Artificial Analysis data shows Gemini 4 Argon ties GPT-6 Astra on the intelligence index while costing about 60% per token.
Nemotron Report: Compatibility in Multi-Teacher On-Policy Distillation
Some teacher models in multi-teacher on-policy distillation (MOPD) are incompatible with each other; the student must be trained on its own rollouts.
Multi-Harness RL: Same Model Performs Differently Across Frameworks
A new deep blog proposes training open models with RL inside real harnesses like Claude Code and Codex to close the performance gap.
Adaption Releases Invent a Dataset Technical Report
Proposes generating trainable datasets from zero data using only a natural-language description.
AI Self-Organization Is Underrated: Math Problem Solved in 88 Hours
Ethan Mollick writes that OpenAI used thousands of agents and about 2.7 million messages to solve a Navier-Stokes-related problem in 88 hours with a very simple coordination structure.
New Benchmark cua-speedrun Reveals Computer-Agent Speed Gaps
Same-score models can differ in speed by up to 4.4x, such as Astra and Kimi K3.
JEV-27B-VL: Open-Weight Multimodal Decision Model
AutoTrustAI calls it the first open-weight, near-SOTA multimodal decision model.
HuggingChat Adds MCP and Automatic Model Selection
ML Intern and Omni modes auto-pick suitable open models per request and connect your own data via MCP.
Perplexity Computer Supports In-Conversation Interactive Charts
Financial data uses TradingView Lightweight Charts for candlesticks, volume and moving averages.
Ideogram 4.5 Launches on ComfyUI
An editing model built to fix drift, keeping the rest of the image intact across edits.
MiniMax H3 Helps Utopai X Rank Second on Video Leaderboard
Utopai X debuted second on the Artificial Analysis text-to-video leaderboard, behind only Wan 3.0.