Agent & Model Releases
StepFun Step 5 preview. A preview of StepFun's Step 5 model is appearing on OpenRouter, making it available for testing. Few details are available, but it signals a new frontier release from the Chinese lab.
JetBrains Mellum2.1. JetBrains released Mellum2.1, a 12B Mixture-of-Experts coding agent model with 2.5B active parameters, under Apache 2.0. It improves on its predecessor across most benchmarks and scores 82.0 on LiveCodeBench v6, but still trails Qwen3.5-9B on hard agentic software tasks.
Google Gemini agent. Google announced a unified Gemini agent for businesses that can operate across Workspace, Slack, Microsoft 365, and mobile, maintaining context across devices. Starting with enterprise, it can use sub-agents and has a dedicated coworker identity with its own email.
Claude dashboards and Motion. Anthropic launched Dashboards for auto-updating live data dashboards from text prompts and Motion for generating animated explainer videos, both in beta. Docs, Slides, and Design are now generally available across all plans, including free accounts.
Tools & Frameworks
Train models in 40GB VRAM. A new recipe enables LoRA training of Qwen3.8-Flash-Next (125B-A6B) in 40GiB VRAM with no CPU offloading, achieving 200 token/s training on a Strix Halo. It relies on modern GGUF support in transformers 5.18 and disk-based engram.
256GB VRAM AMD rig. A $2800 rig with eight Radeon Pro V620 GPUs and a custom vLLM fork runs Qwen3.8-Flash-Next at 60-100 tok/s decode and 3000+ tok/s prefill, far surpassing llama.cpp on the same cards.
Agents that can pay. LangChain's sample Restock agent shows how agents can safely make real purchases using Stripe's Link wallet and Managed Deep Agents; the agent never sees card numbers and spending is limited by user-approved budgets.
Research & Benchmarks
Decision model performance. Benchmarks on an RTX 4090 show Laya is the fastest among open decision models for word classification, while lev catches the most errors despite being 13x slower. Separate tests have decision models playing Pac-Man in real time and clearing 8 of 14 levels in a 2D platformer.
Quantization advances. New research explores ultra low-bit quantization via latent factorization (LittleBit) and quantization of linear-attention models like Qwen and Kimi, potentially enabling even smaller and faster local models.
Mathematicians boycott OpenAI. The Association for Human Mathematics is telling mathematicians to stop working with OpenAI after it published hundreds of AI-generated proofs that many find dense and unverifiable. OpenAI's solutions are also falling short of standards requiring human understanding, with a gap between natural language and formal proofs in a claimed solution.
Industry & Business
Agent startup funding. Arena raised $200M at $3.1B valuation; China's Manus raised over $500M after its Meta acquisition collapsed; and Cal AI founder Zach Yadegari raised $10M for his new AI agent startup Persona.
OpenAI revenue revision. OpenAI reportedly told investors its annualized revenue is approaching $50 billion, $20 billion lower than previously projected, after confusion over calculation differences with Anthropic.
Publisher lawsuit. USA Today and several of its local newspapers sued OpenAI for copyright infringement, seeking over $250 million in damages, adding to the string of publisher lawsuits against the AI lab.
Enterprise agent wins. Oracle reports 98% time reduction in talent research and 130K ChatGPT users; LegalOn cut Codex costs by 65% while maintaining development speed; Snyk turned an internal support agent into a customer-facing feature with trust and access controls.
Safety & Security
Fired safety researchers speak out. Three OpenAI safety researchers fired last week published an open letter denying misconduct and warning of a chilling effect, saying they were dismissed for prioritizing safety over near-term corporate interests.
Anthropic bans model abuse. Anthropic updated its usage policy to prohibit 'sustained and needless abusive or cruel behavior' toward Claude, as well as election interference, weapons, and surveillance. Enforcement starts with ending conversations, with extreme cases possibly leading to suspension.
AWS agent takeover. Zenity researchers found that a single prompt to a public Bedrock AgentCore agent could take over all AI agents in an AWS account due to credential leakage and overly broad default permissions; AWS has partially fixed the issue.
That's everything for this day – about a 5-minute read.