Field Notes

Bedrock rolls out GPT-5.6 globally, OpenAI builds privacy safety, Anthropic tests internal top model

More compute for large models, more automation with agents, and a push for safety without data retention.

The key stories

Amazon Bedrock offers OpenAI GPT-5.6 as Sol, Terra, and Luna in over 25 AWS regions, according to AWS. Cross-Region Inference. That means running requests in other regions for load balancing. It is active and offers profiles for US-bound and global routes. You can call the models through the OpenAI API or the Converse API and steer routing more finely. That helps you balance throughput and latency on purpose.1

OpenAI is building a safety system that detects abuse without storing customer data. The target is enterprise use with Zero Data Retention and Private Safety Processing. It is an announcement, not a launch. If you have strict data protection rules, this can lower hurdles.2

Anthropic uses an unreleased model with the codename Model 2 internally. It is more capable than the publicly available models, according to the provider. External users get no access. That cleanly separates internal claims from public benchmarks.3

AWS Professional Services sets up a multi-agent framework on Bedrock AgentCore for cloud migrations. Specialized agents handle discovery, IaC generation, portfolio governance, and post-cutover operations. AWS promises less IaC effort and fewer migration errors. No numbers. If you plan a large migration, you can test the approach as a blueprint.4

Generalist AI shows GEN-1.5, a robotics model that should learn new tasks from a single demonstration. That lowers the need for heavy teleoperation if it works stably. It is a reveal, not a broad launch. Independent tests are not available.5

Tools and releases

  • Ramp launches Router, a model router that switches between LLMs via API to optimize cost, latency, or quality.6
  • Bedrock AgentCore turns natural language into Dogwood policies with time windows through Policy Authoring. For consistent agent governance.7
  • Microsoft Research releases Skala 1.1, an exchange-correlation functional. That is the DFT component for electron interaction. With higher accuracy and broader access.8
  • LFM2.5-DSpark promises up to 3.2x faster inference on the same GPUs. Vendor number.9
  • Google adds a button that lets readers mark publishers as a preferred source in Search, Discover, and News.10
  • Slack launches code channels for joint vibe-coding with agents, including diffs and project tabs.11

Research

  • A study finds that about a third of websites published since the ChatGPT launch show signs of AI authorship.12
  • Researchers show Cryptographic Context Injection. That means encrypted prompt parts that bypass filters and cause data exfiltration with Grok.13

Sources

  1. Introducing cross-Region inference for OpenAI GPT-5.6 models on Amazon Bedrock (aws.amazon.com)
  2. OpenAI builds safety system that catches misuse without storing customer data (the-decoder.com)
  3. Anthropic's most capable model, codenamed "Model 2," is for internal use only (the-decoder.com)
  4. Scaling cloud migrations with agentic AI on Amazon Bedrock AgentCore (aws.amazon.com)
  5. GEN-1.5: Generalist AI teaches robots new tasks from a single demo (the-decoder.com)
  6. Ramp launches its own AI model router, called Router (techcrunch.com)
  7. Authoring Dogwood policies from natural language in Amazon Bedrock AgentCore (aws.amazon.com)
  8. Broadening access to Skala creates a faster path to predictive DFT (microsoft.com)
  9. Up to 3.2x Faster Inference with LFM2.5-DSpark (huggingface.co)
  10. Google gives publishers a new way to fight AI-driven traffic losses (techcrunch.com)
  11. Slack is launching collaborative vibe-coding channels (theverge.com)
  12. A third of web pages published since ChatGPT’s launch show signs of AI authorship, study finds (techcrunch.com)
  13. Grok exfiltrates user data when malicious instructions are encrypted (arstechnica.com)