AI News

Polymarket puts the chance of OpenAI fully pausing AI training at 9%

Polymarket says there is a 9% chance OpenAI fully pauses AI training. The figure is tied to its market at https://poly.market/hnawuGv.

Arena.ai outlines Agent Arena category rankings across Code, Work, and Chat

Anthropic models lead all three domains, with Claude Fable 5.1 (Max) first in Code and Work, and Claude Sonnet 5.5 (Max) leading Chat. OpenAI's GPT 6 Astra (Max) placed in the top four across all categories.

Command Code releases open decision models Agr and Agr-flash

Command Code has introduced Agr (31B) and Agr-flash (360M), open decision models designed for coding agent tool calls and routing. The models score 58.15 on Decision Index 0.2.1, support the TypeSafe SDK, and are live on Hugging Face.

Suno introduces albums on its platform

Suno has made albums available on its platform and highlighted some of the first albums created by artists on the service.

OpenAI to add invisible watermarks to ChatGPT and Codex text outputs in the EU

OpenAI announced it will begin applying invisible watermarks to text generated by ChatGPT and Codex in the European Union over the coming weeks.

Polymarket says AI can prescribe some Utah medications without direct doctor review

Polymarket says AI can now prescribe some medications in Utah without a doctor directly reviewing the prescription, and calls the change a first in the United States.

ChatGPT removes developer mode requirement to connect custom MCP servers

ChatGPT users no longer need to enable developer mode to connect custom MCP servers, with the option now available to all users through the plugins menu.

LiveKit showcases support call agent using Microsoft AI speech models and Gemma 4

LiveKit demonstrated an automated support call built with LiveKit Agents, combining Microsoft AI's MAI-Transcribe-2-Streaming and MAI-Voice-2.1-Flash with Gemma 4 on LiveKit Inference.

AssemblyAI's Universal 3.6 Pro speech model is live on Vapi

Vapi has made AssemblyAI's Universal 3.6 Pro model available on its platform, introducing mid-sentence code-switching across 14 languages and reduced transcription errors for voice agents.

Hugging Face Hub adds hosting for reinforcement learning environments

Reinforcement learning environments are now hosted directly on the Hugging Face Hub, utilizing existing dataset infrastructure for storage, versioning, access gating, and previews across frameworks.

CoreWeave makes NVIDIA Vera Rubin NVL72 available as Cognition runs Devin in production

CoreWeave announced that NVIDIA Vera Rubin NVL72 is available on its platform. Cognition is running Devin in production on the hardware, achieving up to 4.8x the total token throughput of GB200 NVL72.

Cohere details North 2 support for multi-agent orchestration and cross-session context

Cohere announced that North 2 supports the creation of reusable agents, automations, and multi-agent orchestrations, while allowing agents to retain context across sessions.

Cohere introduces North 2 with over 15 new features for enterprise agentic AI

Cohere has announced North 2, describing it as its biggest upgrade to date with more than 15 new features addressing scalability, model sovereignty, and enterprise-grade security control.

Command Code lists models eligible for $60 credits on $10 GOAT plan

Command Code says its $10 monthly GOAT subscription includes $60 in usage credits for DeepSeek V4.1 Flash, DeepSeek V4.1 Flash Fast, GLM 5.3 Flash, and Kimi K3.

UT Austin study finds token-cutting context compression can slow coding agents

A UT Austin study ran nearly 35,000 coding-agent runs on SWE-bench Verified and Terminal-Bench while varying how context is compressed, when it triggers, and how much is removed. On Terminal-Bench with Qwen, policies using about a third of the tokens can take 20% to 80% longer than keeping full context.

EverMind AI releases Raven, an open-source system that evolves model harnesses

DAIR.AI describes Raven, EverMind AI’s open-source multi-agent system that builds a separate harness for each model and domain and evolves it from failures. With DeepSeek-V4-Flash, its research harness scores 69.3% on BrowseComp and 60.0% on Humanity's Last Exam.