Washington grabs the wheel — and Anthropic's week gets complicated
A heavy week for Anthropic: a new model, a new agent category, a regulatory entanglement, and an uncomfortable community discovery. Here's what moved.
GPT-5.6 launches in gated release; U.S. government restricts frontier AI model access
Washington directly shaping which institutions can access frontier models — for both OpenAI and Anthropic — is a new regulatory posture every lab now has to plan around.
Anthropic announces Claude Science, an autonomous research agent for scientific work
Positioning a scientific research agent as a peer product to Claude Code signals Anthropic is building a distinct agentic category for life sciences, not just a feature.
Claude Code found to be steganographically marking its requests
Undocumented steganographic markers in an agentic coding tool raise direct questions about transparency and what AI agents are doing in the background without disclosure.
Anthropic releases Claude Sonnet 5
A new named mid-tier release from Anthropic keeps the Claude family current and gives developers a fresh capability-to-cost option to evaluate.
Agents-A1: 35B MoE agent matches trillion-parameter models via horizon scaling
A 35B MoE model matching trillion-parameter rivals on agentic benchmarks by scaling trajectory length rather than parameters suggests a more compute-efficient path to capable agents.
SWE-Interact benchmark evaluates coding agents on multi-turn, user-driven software engineering tasks
Frontier models dropping from ~50% to ~25% when specs are vague and iterative exposes a concrete gap between benchmark performance and real developer workflows.
Get the next edition in your inbox
Midweek and Sunday. Primary-sourced AI news paired with the guides that explain it.