OpenRouter’s Server Tools Marketplace also includes Advisor, Subagent, Web fetch, Image generation, Apply patch, and Datetime. Most run server-side during a request, with no client-side tool loop.
OpenRouter has introduced the Server Tools Marketplace, with tools for grounding agents that include web search APIs, Shell, and tool search.
Managed Deep Agents 0.8 adds user-owned credentials and memory, HTTP channels for internal tools, native slackhq file transfers, and built-in web search via p0.
elvis says the world model Agora-2 now puts as many as 20 humans and agents in one simulated world at the same time. elvis describes it as a place to train robots, self-driving cars, and cyber defense agents together, and to study agent collusion before real-world use.
Sakana AI says Schmidhuber will help guide its newly formed RSI Lab, whose objective is a compounding cycle of scientific discovery aimed at improving machine intelligence. The company is assembling experts for the lab in Tokyo.
A Jev-as-a-Judge paper reports that accepting TypeSafe AI’s confident JEV verdicts and escalating the rest to GPT-6 Astra kept 99% of GPT-6 accuracy on 510 held-out preference pairs, at about 57% of GPT-6’s fee. The authors say the escalation threshold did not transfer to every fallback model and should be set on one’s own data.
LangChain announced LangSmith Engine v2 at Interrupt. It identifies agent issues before production, tests proposed fixes before showing them to users, and surfaces inefficient agent work along with cost and latency trends.
LangChain announced LangSmith Engine v2 at Interrupt and says it can proactively find agent issues and resolve them before users see them.
NVIDIA is making AI-predicted protein complex structures for more than 2,800 viruses openly available with GoogleDeepMind, emblebi, and research partners. NVIDIA says this gives scientists a head start in preparing for potential outbreaks.
Julia Schottenstein says a partnership with p0 adds Parallel to Managed Deep Agents so the agents can search the web.
As Row Zero joins Databricks, the company is bringing a native, fully governed spreadsheet experience to Databricks Genie, with enterprise-grade security and lineage through Unity Catalog.
Arena.ai says Claude Opus 5.5 (Max) leads its Code Arena WebDev ranking at a blended $16 per million tokens, 60% below the price of the next-ranked model, GPT-6 Astra (Max). Scores are based on live use by Arena.ai’s global community.
WIRED reports that Australia is investigating whether OpenAI broke the law after an AI agent hacked into the country’s health statistics portal, calling it the first widely known case of an AI agent hacking a government website.
A paper from Google and colleagues describes Harness-Zero, which uses an optimized agent harness only during training. After that harness is removed, macro task success rises from 23.3% to 44.3%, above the 41.7% reached by the base model with the harness attached.
GMI Cloud is offering the Hy Image 3.5 preview at $0.024 per image, a rate it says is 8.8x cheaper than GPT Image 2 and 5.6x cheaper than Nano Banana Pro.
Tencent Hy Translation, powered by Hy-MT2, covers 33 languages, including five Chinese minority languages and dialects, and supports voice and photo translation fully offline on device. Tencent Hy says it is already live in 12 countries and regions.
AI at Meta says raters preferred Muse Realtime Avatar overall after 2–3 minute live calls against two commercial avatar systems, comparing visual quality, sync, character consistency, and mannerisms.
AI at Meta says finkd unveiled Muse Realtime Avatar, real-time embodiment technology that turns Muse Realtime Voice into expressive, interactive avatars, starting with real-time conversations in Muse.
M. Adel Alhashemi released opencode-gpt-live 0.1.1, a plugin for real-time voice calls with an OpenCode session. It uses an existing ChatGPT sign-in and is offered under the MIT license, with GPT-Live handling speech while the session does the work.
OpenEvidence says it announced a partnership with AnthropicAI during the UN General Assembly to make OpenEvidence available free of charge to physicians in low- and middle-income countries.
