Gemini 4 Argon: Google's next frontier model lands
Built for complex workflows across coding, enterprise knowledge work and cybersecurity defense, the model rolls out today to trusted testers under the Fairwind Program.
Google DeepMind has unveiled Gemini 4 Argon, its newest frontier model, and is positioning it squarely at the hardest, most consequential work: software engineering, enterprise knowledge tasks, and cyber defense. Rather than a broad consumer launch, the rollout begins with a set of trusted testers through the Fairwind Program, with wider availability to follow. The sequencing signals a deliberate bet on reliability and safety in high-stakes environments, where a single mistake can cascade across an entire organization and where government and defense partners are expected to be among the first in line.
Runway's Praxis-1 turns video pretraining into robot control
An open-weight World Action Model, Praxis-1 extends Runway's video expertise into real-world robot control. Robot demonstration data is scarce, but video is effectively infinite, so the team bets the best policy models will learn from video, with lightweight fine-tuning to adapt to new robot bodies.
Ideogram 4.5 claims the most precise editing yet
With each edit, leading models add artifacts, pixel shifts and color drift. Ideogram 4.5 is built to eliminate that buildup, making true multi-turn editing possible. It ships in Ideogram, the API and partner platforms, with open weights coming soon.
Biosecurity is one of the most urgent challenges for the AI era.
Pichai gives Argon an early, confident preview
Google's CEO says Gemini 4 Argon shows frontier performance in complex workflows, cyber defense and software engineering, and notes that internal Google teams are already using it extensively across the company.
Cohere ships Embed 5 for frontier and low-latency workloads
The new family pairs Embed 5 Pro for frontier retrieval capabilities with Embed 5 Fast for low-latency performance, which Cohere calls its new state-of-the-art embedding lineup.
Solaris: an operating system that builds itself as you use it
Runway CTO Kamil Sindi calls the first World Interface Model a new kind of operating system. Every frame is generated in real time in response to your intention, so the interface assembles itself around the way you interact with it, rather than being designed once and fixed in place.
Perplexity open-sources a contextual embedding model
pplx-embed-v2-context-9b-preview encodes each chunk with the whole document in view, setting new state of the art on ConTEB and turbopuffer's context-bench.
Meta proposes Context Language Models
CLM treats context as an editable file rather than an append-only conversation, lifting scores by 65% at equal compute on a 24-hour multi-repo agent-swarm task.
Argon opens first to governments and cyber defenders
DeepMind's CEO calls the launch a key capability leap, rolling out first to government and trusted cybersecurity defenders before wider availability.
Runway Ads puts autonomous agents on performance marketing
The engine generates video and image ads from a brand's asset library, publishes approved versions to Meta, Google, TikTok and Snap, then reads back actual spend data to drive the next round of creative. Early access is open.
Ollama now runs Jev-style decision models locally
A new local /v1/systemone API can pull decision models such as Nimble for ticket triaging, model routing and content moderation, bringing fast decision systems fully on-device.
Anthropic opens a home for Claude developers
The new claude.dev site gathers engineering deep dives, Claude Code and API guides, and team tips, alongside a few Easter eggs.
Tencent releases ExplorationBench
A benchmark for measuring how AI systems explore, framing hypotheses, designing experiments, and learning from results.
SGLang turns Qwen3.8-27B into a decision model
Its native /v1/decisions API converts LLMs and VLMs into classifiers, and the multimodal build clears Pokémon FireRed with sub-100ms decisions.
Letting LLMs edit their own context improves memory
Research finds giving models direct control over context yields better long-horizon memory and FLOP-efficient adaptation.
DSPy's Flex lets optimizers rewrite the module itself
Flex is code rather than a fixed prompt; GEPA can split a task into predictors plus ordinary Python, executed in a sandbox.
Stripe acquires OpenRouter and its token economy
A Latent Space discussion unpacks the model-routing deal, the ten trillion tokens processed daily, and the trillion-token economy ahead.
Figure decommissions its robot in a vat of molten steel
The second-gen Figure 02 is retired by melting it down in a Finnish foundry; the metal becomes limited-edition souvenirs.
Argon agents optimize data-center memory
A team of Argon agents reportedly analyzed fleet-wide telemetry and applied memory optimizations, freeing over 300 TiB with up to 1 PiB projected.
DeepSeek's kernel work makes Ascend viable for training
Commentators say the Ascend ecosystem is now feasible for end-to-end training, recalling Liang Wenfeng's line that someone has to step onto the frontier.
OpenAI maps how small businesses use AI agents
A new report explores agents for customer acquisition, product work and finance, alongside ASBDC training.
ChatGPT adds shareable Sites and plugin profiles
Shareable profiles let others find and reuse what you have built.
Hedra ships MCP and CLI for vision models
Connect Claude Code, ChatGPT or custom agents to Hedra's image, video and world models.
Recraft V4.1 Flash renders in 1.3 seconds
Prompt to image in 1.3s; V4.1 Pro refines at 2048×2048 while preserving composition and lighting.
Perplexity Computer takes tasks by email
No account required, forward or CC an email and the agent completes tasks in the background.
Qwen3.8-27B arrives on Nebius
The 27B dense model is now callable for agent building and multi-step deep research.
HeyGen Video builds on MiniMax H3
The new video product is post-trained on H3 to bring production quality at lower cost.
Creatify debuts Boreal-H3 ad model
Optimized for ads, it follows creative briefs while keeping product and character consistent.
Replit apps run inside Meta VR
Users can now run apps built with Replit directly on Meta VR.
ElevenLabs v4 lands on Higgsfield
Add whispers, laughter and sound effects in scripts across 90+ languages.
Kling shows off Kling 4.0
A full-capability demo video highlights upgraded motion and image quality.
NVIDIA proposes one-shot video distillation
LongLive-Plug learns few-step sampling and long-context correction as LoRA, plug-and-play.
Omni-IO Skills give agents multimodality
27 skills across 7 modalities lift host input support from about 40% to 100%.
LEGO-Anything rebuilds 3D scenes with code
A coding agent writes and runs Blender scripts; best models hit 53.4% indoor reconstruction.
Audio8 ASR Infinite streams 24/7
Native streaming decodes 12.5 times per second with constant memory and latency.
Vercel Connect simplifies agent integrations
A safer, simpler way for agents and apps to connect to services.
Cursor visits 7-person robot team Innate
The co-founder says seven people are doing the work of fifty on general-purpose robots.
Vidu partners with Reactor on S2 Avatar
The tie-up brings S2 Avatar to more global creators.
Luma batch-generates ad variants
One campaign can produce multiple sizes for different markets.
Interactive avatars make articles answerable
Readers can question a reporter's digital twin for more background.