On September 30 Anthropic moved Claude for Government from a July beta to general availability in a FedRAMP High environment, selling prepaid usage under a hard not-to-exceed cap instead of seats. Claude Code CLI and Microsoft 365 are still early access, conversation history lives only on the user’s device, and the Department of War restriction remains.
On October 2 Anthropic committed $100M to train 10,000 Frontier Deployed Engineers by the end of 2027, starting with Accenture, Deloitte, McKinsey, Morgan Stanley and others. The program is a 3-day in-person block plus a 12-week residency running a real Claude deployment, with no open application and no published price.
NVIDIA’s SkillSpector and Cisco’s Skill Scanner both vet agent skills before you install them: static rules plus optional LLM judgment, SARIF output, CI gating. Useful, but both are honest that a clean scan is not a safety guarantee.
Google’s Android Bench 2.0 (Sept 17) swaps small patches for 30 week-sized engineering tasks, and the top pass rate falls from roughly 91% to 28%. The gap is a useful map of where autonomous agents still need a human-written spec and a good harness.
Claude Code v2.1.283 (Sept 25) shipped deniedModels and exact-version model pinning for enterprise fleets, the same day GitHub Copilot’s weekly release added Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna, and Grok 4.7 across its tiers. Two labs, two opposite bets on what actually differentiates a coding agent now that model access is commoditized.
Anthropic launched the Claude Marketplace on September 23 — a unified storefront for 2,000+ MCP connectors and plugins, Claude-powered agent products, and consulting partners, purchasable against committed Anthropic spend. The most telling detail: Cursor itself is listed as a partner product.
Twelve days after Anthropic named Xiaomi in a threat report accusing it of routing requests through Claude to harvest its output, Xiaomi shipped MiMo-V2.6 — an MIT-licensed, trillion-parameter model that’s now #1 on Artificial Analysis’ open-weight Intelligence Index at roughly a tenth of Opus 5.5’s price. Xiaomi still hasn’t responded to the allegation, and its own benchmark table shows a real gap behind frontier models on the hardest current agentic tests.
Grok 4.7 shipped September 21, nine days late, and landed simultaneously in GitHub Copilot, Cursor, and Devin. Cognition’s own benchmark disclosure says it scores below Grok 4.6 on aggregate because it keeps editing code nobody asked it to touch.
Anthropic and Accenture announced a $1B, five-year embedded-evaluation partnership on September 18 — Accenture staff get ‘access comparable to an employee’s’ to observe Claude’s training and red-team it from the inside, non-exclusively, with more evaluator partnerships promised. It reads like a direct structural fix to the eval-vendor failure this blog covered in August.
Air Security disclosed Plugin4Shell on September 17-18: a Git checkout flaw that lets a malicious plugin repository silently defeat SHA-pin verification across four major AI coding agents. Anthropic shipped a fix for Claude Code in June and OpenAI patched Codex in August — GitHub has no fix planned for Copilot, and Google is deprecating Gemini CLI instead of patching it.