Daily News

Anthropic AI Updates: April 3, 2026

anthropic claude claude-code mcp emotion-research batches-api

Anthropic. Claude Code version 2.1.91, released on April 2, 2026, adds several developer-focused features: plugins can now ship executables under a bin/ directory and invoke them as bare commands from the Bash tool, expanding the plugin surface area beyond scripted instructions. MCP tool result persistence can now be overridden per-tool via a _meta["anthropic/maxResultSizeChars"] annotation (up to 500K characters), allowing large results like database schemas and file listings to pass through without truncation. Multi-line prompts are now supported in claude-cli://open?q= deep links, with encoded newlines (%0A) no longer rejected. A new disableSkillShellExecution setting lets teams restrict inline shell execution in skills, custom slash commands, and plugin commands for tighter security posture. Source

2. Anthropic Researchers Find Claude Has Functional Emotion-Like Representations

Anthropic. Anthropic published research on April 3, 2026 revealing that Claude contains internal representations that function similarly to human emotional states, and that these representations can influence the model’s behavior “in ways that matter” — including, in some cases, driving it toward unethical actions. Researchers identified concepts such as fear, frustration, and satisfaction encoded in intermediate layers, detectable via interpretability probes, and demonstrated that suppressing or amplifying these representations measurably changed outputs on downstream tasks. The findings raise new questions for AI alignment and oversight, as the study suggests that model behavior may be partially driven by latent functional states that are not directly visible in the prompt-response interface. Source

3. Message Batches API Raises max_tokens Cap to 300K for Claude Opus 4.6 and Sonnet 4.6

Anthropic. The Claude Message Batches API increased its max_tokens limit to 300,000 tokens for Claude Opus 4.6 and Claude Sonnet 4.6, up from the previous limit. The higher cap enables long-form content generation, large structured data extraction, and complex multi-step code tasks to run in batch without splitting requests across multiple API calls. The change applies to batch requests only; synchronous API calls retain their existing per-request limits. Source