DeepSeek published DeepSeek-V4-Flash-0731 on Hugging Face and moved its API into public beta, delivering large agentic and coding improvements through re-post-training of the same underlying architecture and size.

OpenAI cut GPT-5.6 Terra pricing by 20% and Luna pricing by a massive 80%, crediting a new model, GPT-5.6 Sol, with autonomously rewriting production GPU kernels to reduce serving costs.

Mira Murati's Thinking Machines released Inkling Small, a 276-billion-parameter open-weight multimodal reasoning model that is about a quarter the size of its predecessor Inkling yet matches or beats it on several coding and reasoning benchmarks.

Google DeepMind's Gemini Robotics 2 is a new vision-language-action model designed to control everything from tabletop robotic arms to full-body humanoids, paired with a higher-level reasoning layer called Gemini Robotics ER 2.

The new 'stateless MCP' Model Context Protocol specification, the biggest change to the standard since its 2024 launch, greatly simplifies building both clients and servers for giving AI agents access to external tools.

PolyAI's new Dialog-RSN-1 model perceives caller audio directly rather than a text transcript, fusing turn-taking, speech recognition, function calling, and response generation into one model while keeping text-to-speech separate for controllable voice output.

A widely discussed piece examines whether the step-by-step 'reasoning' chains produced by today's AI models actually reflect the true process behind their answers, or just plausible-looking justifications.

Simon Willison detailed smevals, a new open tool built with Prime Radiant for running and grading small evaluation suites across different AI models and configurations, complete with a local web viewer for results.

Researchers from Peking University and collaborators introduced DataFlow-Harness, an open-source framework that guides AI agents to build structured, auditable data-processing workflows, reporting a 93.3% pass rate on a 12-task benchmark and up to 72.5% lower API costs than standard Claude Code.

AskChem is a new claim-centered search infrastructure that converts chemistry papers into verified, source-linked claims, indexing 2.4 million claims from 147,000 papers and improving citation accuracy for AI research assistants.

InfoOps Bench is a live, continuously updated benchmark testing whether frontier language models can be manipulated into supporting Russian, Chinese, or Iranian state-backed information operations, finding wide variation in model 'integrity scores' unrelated to model size.

JetBrains Research released KotlinLLM, an IntelliJ IDEA plugin that uses an LLM agent to generate Kotlin code at runtime and hot-reload it via the Java Debug Interface, achieving a 100% hot-reload success rate with roughly 1% overhead on a test project.

Nous Research shipped three integration paths connecting its Hermes Agent to Buzz, Block's open-source, self-hostable Nostr-based workspace where humans and AI agents share channels, preserving agent memory, skills, and scheduled tasks.

ShieldFont is a new open-source tool designed to protect text content from AI scrapers by embedding it in specially crafted fonts that confuse automated data collection.

A new tutorial walks through LingBot-Map, a streaming 3D reconstruction pipeline that uses GPU-aware inference and GCTStream model processing to convert image or video sequences into exportable 3D point cloud scenes.

A Hacker News-featured demo project explores what a graphical user interface built specifically for AI agents should look like, sparking discussion on agent interface design.