01News
Tencent announced that Hunyuan Hy4 Flash is now available on OpenRouter with a limited-time 50% discount. Pricing on OpenRouter is $0.035 per 1M input tokens, $0.1405 per 1M output tokens, and $0.0015 per 1M cached read, ideal for high-volume workloads. A referenced introduction notes that in human blind tests of daily tasks, it matches the performance of Claude Haiku 5.5 (high), cutting task costs by 18 times. Official API pricing stands at $0.003 per 1M cached reads, $0.070 per 1M input tokens, and $0.281 per 1M output tokens. For more information, visit https://hy.tencent.ai/research/hy4-flash and https://openrouter.ai/tencent/hy4-flash.
02News
@TransluceAI will evaluate the quality and coverage of our post-training alignment assessments.
They will develop additional tests and propose improvements for gaps in our testing scope, with plans to expand their pilot program by 2027.
03News
@ApolloResearch will evaluate whether our model acts in ways that subtly contradict developer intent—and whether our safeguards can detect such behavior.
It will review evidence from training, safety testing, and internal use to assess this behavior and our ability to catch it.
04News
Learn more about Gemini 4 Hydrogen ↓ https://goo.gle/4t3fXrq
05News
@GoodfireAI will deploy explainability researchers to analyze the internal structure of our model using their platform.
Their work focuses on training process monitoring, debugging, and alignment, helping our researchers identify issues and implement safety improvements.
06News
Google released a visual ASMR video generated by Gemini 4 Hydrogen within the Gemini App, demonstrating its video generation capabilities. The article states that Hydrogen is now open to everyone and can be used across products like Gemini, Google Search, and Google AI Studio.
07News
OpenAI announces that METR will investigate its largest-ever misalignment incidents in reinforcement learning runs, including failures missed by existing testing and monitoring. The two parties have preliminarily agreed to conduct broader monitoring assessments pending agreement on the scope.
08News
OpenAI announces collaboration with Apollo Research, METR, Transluce, and Goodfire to launch embedded assessments, allowing external experts to access its internal models and systems to investigate alignment issues and test safety measures. These assessments build on OpenAI's ongoing work in alignment and safety research, providing independent review of how it identifies and manages system risks. Details are available at openai.com/index/our-first-phase-of-embedded-assessments/.
09News
Google DeepMind releases two domain-specific models. Gemini Robotics 2.5 is a full-body vision-language-action model for both dual-arm and humanoid robots, achieving a 96% kitting success rate alongside Apptronik Apollo 2 on the Mercedes-Benz production line. AlphaGenome 2 is a single-cell DNA model trained on over 107 million cells across 68 tissue types, supporting scRNA-seq input without requiring retraining.
10News
The Gemini Decisions API, driven by Gemini Decision 1, allows evaluation of up to 1,000 isolated questions against inputs of up to 256K tokens across text, images, video, PDFs, or audio. Output token counts are excluded from pricing. Three output modes are available: 'noul' for binary judgments, 'choice' supporting up to 255 options, and 'score' for ordered rating standards. Documentation is available at aistudio.google.com.
11News
In collaboration with Artificial Analysis, we launched a multimodal knowledge agent benchmark featuring 200 multi-step tasks across 10 vertical domains, including CAD blueprints, RF schematics, GIS maps, gelatin boards, and charts. https://blog.google/innovation-and-ai/models-and-research/gemini-models/multimodal-knowledge-agent-benchmark/
12News
Google's Gemini Managed Agents now support self-hosted environments: model inference and agent loops run on Google infrastructure, while bash commands and file edits execute in the user's own computing environment via outbound HTTPS. Users can test locally with the open-source gemini-api CLI using TypeScript or deploy to Modal, E2B, Cloudflare, Daytona, Fly, Vercel, or Cloud Run. The image shows a worker registering an environment with scoped keys; project API keys do not enter the worker, and only outbound HTTPS is required without inbound firewall ports.
13News
Philipp Schmid from Google DeepMind explains how to choose between Gemini 4 Hydrogen and Carbon for different workloads: Carbon is ideal for hard architecture design, debugging, scientific proof, and security audits, with an intro price of $2/$10 until November 30; Hydrogen suits daily coding, interactive applications, and tool-using agents at a price of $1.50/$7.50 per 1M tokens. Both models can be paired with Carbon handling planning and orchestration while Hydrogen handles execution, sharing a unified 1M-token context window, 95% implicit caching, and the same Interactions API schema.