Google has launched a decision-making API based on Gemini Decision 1 for tasks such as routing, tool selection, and safety checks. The model accepts text, image, audio, video, and PDF inputs; a single request can contain up to 1,000 yes/no, multiple-choice, or rating questions, and returns answers with confidence scores. The API supports a 256K context window. Input costs $0.10 per million tokens, while output is free. The model is also available on OpenRouter.

Anthropic has launched its IPO roadshow and plans to offer 825 million shares at $70 to $78 each. Roadshow materials show that the company generated $16.956 billion in revenue in the third quarter of 2026, up more than 12-fold year over year, with annualized revenue of about $81 billion. Its adjusted operating margin was 15%, with a long-term target of more than 35%.

Google DeepMind has released AlphaGenome 2, a DNA sequence model designed to predict how different cell types interpret gene sequences and how DNA variants affect those cells, helping researchers explore links between genetic variation and disease mechanisms. The model is available for non-commercial research through an API and an interactive Colab tutorial.

Baidu has launched ERNIE 5.5 Flash Preview for coding and agent tasks, and opened it for testing on Arena. The model has 280 billion total parameters and activates 14 billion parameters per token, with support for a 1-million-token context window. In demonstrations, it wrote C++ code continuously for 11 hours, built a game using Unreal Engine, and handled errors automatically. It scored higher than Claude Opus 5.5 on AutomationBench and SpreadsheetBench V2.

Runway has released Praxis-1, a world-action model for robotics that uses video pretraining to support physical robot control. Its weights are available under the Apache 2.0 license, allowing researchers and developers to run and fine-tune the model on their own hardware.

Alibaba plans to spin off its in-house chip design division, T-Head, and pursue an independent listing on the Hong Kong Stock Exchange. It has submitted an application to the exchange. T-Head is Alibaba's chip design division.

vLLM has published inference benchmark results for MiniMax-M3.1-Flash on NVIDIA Vera Rubin NVL72: total token output, including cached tokens, reached 10.2 times that of the GB300 NVL72 configuration. Under low-concurrency conditions, the 90th-percentile per-user generation rate peaked at 1,205 tokens per second.

GitHub Agentic Workflows are now generally available, allowing developers to run coding agents in GitHub Actions for tasks such as issue triage and continuous integration failure analysis. Workflows are configured using Markdown and YAML and run in sandboxed containers protected by a firewall. Agents are read-only by default; write operations are handled by separate tasks, and approval gates can be configured. The default cost cap is $10 per run, while enterprises can set budgets and timeouts.

The Institute of Foundation Models at Mohamed bin Zayed University of Artificial Intelligence has released PAN 2, an open-source world model for simulating physical environments and supporting robot planning. The model consists of a 1.5-billion-parameter language model and a 29-billion-parameter implicit diffusion model. It can predict the outcomes of physical interactions and maintain spatial consistency across simulated time steps.

Perplexity has published results from its SPACE virtual machine sandbox red-team test, in which 12 AI models were asked to escape a VM to obtain host credentials or bypass network policies to access blocked URLs. The test found that prompts based on traditional capture-the-flag tasks could lead agents away from the actual security boundaries. Models fine-tuned specifically for cybersecurity did not necessarily outperform general-purpose models.