Kev is a LoRA adapter and pointer head on Qwen3.5-4B-Base
OpenRouter says Kev is a LoRA adapter and pointer head on Qwen3.5-4B-Base, with weights in a Hugging Face repository, and thanked SiliconFlowAI for hosting.
AI News
Kev uses the same OpenRouter API as Jev
OpenRouter says developers already calling Jev can switch to Kev by changing only the model field to jaredpalmer/kev-4b.
Kev 4B is now available on OpenRouter
jaredpalmer’s small open-weight System One model accepts state and typed questions and returns typed answers. OpenRouter prices it at $0.042 per million input tokens and $0 for output, with an 8K context window.
Salesforce makes Agent Optimizer available in beta
Agent Optimizer can analyze production sessions, spot failure patterns, investigate and test a fix, and stage changes for review. An autonomy dial sets how much of that work it does on its own.
Grok iOS app adds multiple SpaceXAI accounts
Evan Bacon says users can now add multiple SpaceXAI accounts in the Grok iOS app and switch or add them from the profile button in the top-left corner.
Google AI releases Gemini 3.8 TTS models, Live Avatar, and Notebook updates
Google AI’s weekly releases add Gemini 3.8 Flash TTS and Flash-Lite TTS, Gemini 3.8 Live with Live Avatar, and Gemini_Notebook tools for all users, including mobile Live Chat in about 100 languages. Project Suncatcher is set to launch a prototype satellite to test Google TPUs in orbit.
Grok-4.7 is now available on WorkBuddy
WorkBuddy has added Grok-4.7 to its model lineup, making the model available for tasks on the platform.
ClaudeDevs says Opus 5.5 is 20% cheaper per token and 60% cheaper on cache reads
ClaudeDevs says Opus 5.5 costs 20% less per input and output token than Opus 5, and 60% less on cache reads. It also published a calculator for estimating Claude Code task cost from /usage.
Perceptron Mk1.5 matches Mk1 pricing on OpenRouter
OpenRouter prices Perceptron Mk1.5 at $0.15 per million input tokens and $1.50 per million output tokens, the same rate as Mk1.
Perceptron Mk1.5 adds audio input, video tracking, and tool calling
OpenRouter says Perceptron Mk1.5 adds audio input, video tracking, and tool calling since Mk1. The model also supports graded reasoning from none to high, structured outputs, and a 36K context window.
Perceptron Mk1.5 is live on OpenRouter
OpenRouter now offers Perceptron Mk1.5, which it describes as Perceptron’s embodied reasoning model for physical agents. The model accepts text, images, video, and audio, and can respond with points, boxes, polygons, and video clips as well as text.
Salesforce introduces Hunter for goals spanning days to months
Salesforce says it introduced Hunter at Dreamforce and describes it as the first agent to run on the company’s long-horizon runtime. Hunter can pursue goals over days, weeks, or months, adjust as new information arrives, and take user feedback while the user stays in control.
skills.sh reaches 1 million agent skills and nearly 280 million installs
Vercel says the skills.sh registry hit 1 million agent skills and nearly 280 million installs within seven months.
Claude completes a nine-loop calculation in planar N=4 super-Yang-Mills
Anthropic says one prompt sent Claude through days of largely unsupervised work in Claude Science, producing a nine-loop scattering-amplitude result that Lance Dixon independently verified, at a total cost of a few thousand dollars.
Runway MCP connects to Claude for image and video generation
Runway MCP can be connected directly to Claude, including Claude Opus 5.5, to generate images and videos with Gen-4.5, Seedance 2.5, GPT Image 2, Kling, and other models.
Arena.ai says Claude Opus 5.5 is easier to read than Opus 5 but writes longer answers
Arena.ai compared high-reasoning Text Arena outputs and says 10 of 12 writing measures moved in a better direction from Claude Opus 5 to Opus 5.5. Average answers grew from 453 to 481 words, while hedges such as “perhaps” rose 97%.
ComfyUI announces 100% Higgsfield API cashback through Comfy Router
Higgsfield announced 100% cashback on API models including Seedance 2.5, Kling 3.0, MiniMax H3, and Wan 3.0, up to $100,000 per business from a $20 million pool. ComfyUI said the cashback also applies through Comfy Router, with unused cashback expiring on September 30.
Proaction is building GPT-Live-1 voice agents for vehicle maintenance
OpenAI Developers says Proaction is using GPT-Live-1 to build voice agents that talk to drivers, call repair shops, and help arrange vehicle maintenance, with the team stepping in when needed.
Replit acquires Atta to add AI analytics to its builds
Replit said on September 25, 2026, that it has acquired Atta to bring AI analytics into its builds. Atta co-founder Omar Shaik said the new analytics experience can be tried in Replit.
SpaceXAI Memphis describes Mid-South sites as large AI training clusters
SpaceXAI Memphis says SpaceX’s supercomputing facilities in the Mid-South rank among the world’s most advanced AI training clusters, spanning more than 2.5 million feet, with millions of GPUs and more than two gigawatts of compute.