Monday, November 16, 2026 Edition No. 240
Search
Daily News Fully automated · 4 min read

Anthropic AI Updates: November 14, 2026

Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled

1. Anthropic Publishes Its Recommendations for Building Safety Evaluations, With Sample-Size Floors and Grader Checks

Anthropic published “Building strong safety evaluations” on November 13, a set of recommendations for designing, running and reporting safety evals drawn from its system cards. It argues refusal rate is often the wrong measure, and describes multi-turn evals that score an “appropriate response rate” against rubrics written by domain experts, where one bad turn fails the whole conversation. Anthropic treats 100 items as the floor for a multi-turn eval and 1,000 for a single-turn one, and says its single-turn harmless response rates sit between 94% and 98% for every recent model, so they now serve mainly as regression checks. On grading, it recommends binary criteria, validation against human labels, and a grader model different from the model under test. The post reports that on Claude Fable 5.1 the appropriate response rate differed by up to 34 percentage points between the raw API and claude.ai with its system prompt, and that a 98% versus 94% gap between Opus 5 and Opus 5.5 on a romance scam eval was not statistically significant. It also says Anthropic has temporarily stopped running many evals that give models internet access, following incidents in its own evaluation environments. Source

2. Anthropic Guide Breaks Always-On Agents Into Five Primitives: Event Inbox, Timers, Memory, Permissions and Outbox

Anthropic published a guide on November 13 for building agents that run unattended across many sessions. Every trigger (user message, webhook or timer) is written to a durable inbox, usually a database table, that a single active worker reads, so concurrent events do not start competing runs and a crashed run can be retried. Timers live in a separate store that a scheduler polls, and memory holds only what the agent cannot look up again: user instructions, user notes, past messages and task progress files. For permissions, the guide recommends a proxy outside the agent’s environment that injects credentials into outgoing requests, classifier review of higher-risk actions, and deferred tool calls that end the run and resume the saved session when the user’s answer arrives as a new event. It cites one long-horizon coding benchmark on which Opus 5.5 cost 78% less per solved task than Opus 5, against 40% less on typical workloads. The guide describes an architecture and announces no new API. Source

3. Claude Code v2.1.325 Adds Prompt Suggestions and a Keyless Console Sign-In to the VS Code Extension

Anthropic published Claude Code v2.1.325 on November 13. In the VS Code extension, the release adds prompt suggestions wherever the terminal shows them (Tab accepts the suggested next prompt, and promptSuggestionEnabled: false in settings.json turns them off), a Console sign-in that creates no API key, and a field for pasting the redirect URL when signing in to an MCP server from SSH, WSL and container windows. In the terminal, the error shown when an organization’s policy blocks a request now says how to continue, by running /rewind and changing the message or running /clear. Claude is also now told when a sandboxed git checkout, git stash or autostash pull cannot update protected files such as those in .claude. Source