The AI log, agents and coding · frontier · science
Covers 11 May 2026 to 17 May 2026
Claude Platform on AWS, Codex on your phone, Grok Build beta
Anthropic's native platform arrives inside AWS accounts, Codex reaches ChatGPT mobile, and xAI opens a coding CLI in beta.
By Ali Asif, published
What shipped
- Anthropic
Claude Platform on AWS
Anthropic's native platform, including Managed Agents in beta, is now available through an AWS account with IAM sign-in and AWS billing, from May 11.
- OpenAI
Codex in the ChatGPT mobile app
From May 14 the ChatGPT mobile app can drive Codex running on a connected Mac.
- xAI
Grok Build beta
xAI's release notes record a coding agent that runs as a terminal app, headless in scripts, or behind the Agent Client Protocol.
- Anthropic
Fast mode for Claude Opus 4.7
A research-preview faster output mode for Opus 4.7 on the Claude API, noted May 12.
- arXiv
Achieving Gold-Medal-Level Olympiad Reasoning via Simple and Unified Scaling
A scaling recipe aimed at olympiad-level reasoning, third on the week's Hugging Face papers list.
- arXiv
MinT: Managed Infrastructure for Training and Serving Millions of LLMs
Infrastructure aimed at training and serving very large numbers of language models.
Thin week for model launches, so most of what I tracked was how agents get delivered.
Claude Platform on AWS
On May 11 Anthropic and AWS announced general availability of Claude Platform on AWS. You reach Anthropic's own platform through your existing AWS account, sign in with IAM, and pay through AWS Marketplace on consumption. Activity shows up in CloudTrail. Features at launch include Managed Agents in beta, the advisor tool, web search and fetch, code execution, the Files API, Skills and the MCP connector. Models at launch were Opus 4.7, Sonnet 4.6 and Haiku 4.5.
The detail to read slowly is data handling. Anthropic operates the service, and customer data is processed outside the AWS security boundary. That is a different arrangement from Bedrock, where AWS runs the inference stack, and the two companies' documents word the data-processor roles a little differently. For regulated buyers, that sentence will matter more than any feature list.
Agents on your phone and in your terminal
OpenAI made Codex usable from the ChatGPT mobile app on May 14, by connecting to a Mac that runs the Codex app. xAI's release notes record Grok Build entering beta in May, a coding agent you can run in an interactive terminal, headless in scripts, or behind the Agent Client Protocol. Anthropic noted a research preview of fast mode for Opus 4.7 on May 12.
Three vendors, one pattern: the agent runs somewhere else, and you check on it from wherever you are. That suits how I work. I am rarely at a desk when something needs approving.
On May 13 Anthropic also put cache diagnostics into public beta. If you have ever watched a long agent run burn tokens because a prompt prefix silently stopped caching, you will see why that is more useful than it sounds.
Papers
A scaling paper on olympiad-level reasoning ranked third in the Hugging Face weekly list. SenseNova-U1 proposed a single architecture for multimodal understanding and generation. MinT described infrastructure for training and serving millions of language models.
What I'd watch next: Google I/O on May 19, since three vendors have now shown a hand and the fourth tends to show up with a model.