Runway adds Layers to split images into editable layers
Runway has introduced Layers, which splits any image into editable layers in one click so users can remove backgrounds, edit text, and rework individual elements without leaving Runway.
AI News
Microsoft paper reports a 16.6-point average gain for CASD prompt optimization
CASD has an off-the-shelf coding agent analyze a full set of agent logs and write the findings as rules in one prompt, with no environment access or validation data. DAIR.AI says one pass beat GEPA and SkillOpt on the reported benchmarks at about $1.60 per optimized prompt.
Cloudflare introduces Turnstile Spin to add server-side verification with an AI coding agent
Turnstile Spin uses a developer’s preferred AI coding agent to wire up the server-side verification missing from incomplete Turnstile setups. Cloudflare says skipping that validation leaves sites exposed to bots.
AWS says Buildsimple improved average data-extraction accuracy from 75% to 90%
Some customers reached 99% accuracy. Amazon Web Services also says Buildsimple has more than 3,500 AI models in production and deploys new AI solutions in 2 hours instead of 12 months.
Polymarket says Grok will soon join X group chats
Polymarket says Grok will soon be added to X group chats and will be able to interact as another member of the conversation. No rollout date is given.
Arena.ai launches redesigned Leaderboard Overview
The new overview gathers live signals on newly added models, category performance, capability notes, and Arena news, and can show cross-arena scores for models such as GPT-6 Sol and Claude Opus 5.5.
Chrome adds practice quizzes and media Q&As with Gemini
Chrome’s new learning features include practice quizzes and media Q&As with Gemini in Chrome, along with picking up where you left off across devices.
LangSmith fine-tuning CLI smithtune trains on FireworksAI_HQ and baseten
Ankush Gola says smithtune pulls, selects, and processes trajectories from LangSmith, trains on FireworksAI_HQ and baseten, and evaluates the results with LangSmith. The code is in the langchain-ai GitHub repository.
Arena.ai ranks GPT-6 Luna (Max) 24th in Code Arena WebDev
GPT-6 Luna (Max) scores 1,593 points, 74 above GPT-5.6 Luna (xHigh) at 41st. Arena.ai says it matches Gemini 3.7 Flash High and Qwen 3.8-27B at a blended $0.40 per million tokens.
Google to test whether TPUs can operate in space
Sundar Pichai says Project Suncatcher will fly a prototype satellite on SpaceX’s Transporter-18 mission, in partnership with Planet, to test whether Google’s TPUs can survive and operate in space.
Scale AI and Google Cloud partner on a reference architecture for the Scale GenAI Portfolio
The joint blueprint brings the Scale GenAI Portfolio to Google Cloud and includes Gemini Enterprise integration. Scale AI says it is meant to give builders a proven foundation for putting AI to work faster.
Databricks introduces the Unity Gateway CLI for coding agents
Administrators can manage models, tools, policies and budgets in Unity Gateway while developers keep using Claude Code, Codex, OpenCode and other coding agents. One configuration update is applied across supported agents.
LangSmith Custom Apps is now generally available
LangChain says LangSmith Custom Apps can create interfaces from agent data with a prompt, including views for labeling runs, comparing experiments, and reviewing traces.
Slack says Slackbot in Lightning can update deals beside a Salesforce record
In Salesforce Lightning, Slackbot can be used next to a Salesforce record to update deals or draft a case resolution without switching tabs.
Recraft V4.1 Flash is now available in ComfyUI
ComfyUI says Recraft V4.1 Flash generates an image in 1.3 seconds and calls it the fastest image model on the market. A selected result can then be refined to 2048×2048 Pro quality.
Fireworks says it is a launch training partner for LangSmith Fine-Tuning and smithtune
Fireworks says it is a launch training partner for LangChain’s LangSmith Fine-Tuning and smithtune, describing a one-CLI path from LangSmith traces to managed SFT on Fireworks, then evaluation and deployment, without separate training infrastructure.
Midjourney releases new --tile and inpainting versions, previews real-time models
The --tile update adds seamless patterns in v8.1 and v8.2, and the new inpainting changes only pixels a user selects. Real-time models are an early preview on Midjourney’s alpha site, not a full release.
Managed Deep Agents 0.8 supports user memories
Managed Deep Agents 0.8 adds user memories with control over where they are stored. Nathan Drezner said memories created in private contexts are never shared, while Context Hub lets developers inspect general agent memories and keeps user memories private.
Hassan trains Tev1 0.8B, a tiny Jev-like classifier
The model runs entirely on his Mac through Ollama and, by his account, classifies tasks in about 50 ms end to end. Weights and benchmarks are not available yet; Hassan says both will be released very soon.
Adithya S K releases SmolDataEnvs with 5,000+ verifiable RL tasks
SmolDataEnvs is an open-source set of more than 5,000 verifiable reinforcement-learning environment tasks for hill-climbing small models in code and data science. The release includes environments, evaluations, and training.