Anthropic AI Updates: August 15, 2026

1. Anthropic Explains How Claude’s Text Watermark Works

Anthropic. Anthropic published a technical explainer describing how its text watermark operates, built on Google DeepMind’s SynthID-Text method. Instead of using a standard random number generator when several words are equally valid, the system uses a cryptographic key to deterministically choose among those options, embedding a statistical pattern that a detector holding the key can later verify without changing how the output reads. Anthropic says the signal is sparse where specific wording is required for accuracy, weaker on heavily edited or rewritten passages, and more reliable on longer text, and it plans to offer a watermark detection API with details still being finalized. Source

2. Anthropic Shares Tactics for Cutting Token Costs in Claude Code

Anthropic. A Claude Code post lays out practical steps for keeping sessions efficient by controlling what enters the context window. Recommendations include running /clear between unrelated tasks, setting the model and effort level at the start because changing them mid-session invalidates the prompt cache, using @-mentions to attach files instead of triggering Read calls, and quieting verbose command output or offloading it to subagents. The post also advises checking /context in fresh sessions to drop unused MCP servers or tools and running /compact before a break, since the prompt cache expires after roughly an hour. Source

3. League Reports Halving Development Cycle Times with Claude

Anthropic. A customer story details how League, which builds consumer health experiences for health plans and providers, deployed Claude Enterprise company-wide in December 2025 across engineering, finance, security, and design. League built an internal orchestration tool called Swarm on Claude Code that runs overnight autonomous coding sessions, and reported cutting development cycle times in half from idea to pull request while engineers merge two to three times more pull requests weekly. The company also said Claude adoption reached 98%, AI-authored code rose to about 98%, vendor security assessments dropped from multiple weeks to 15 minutes, and it automated more than 60 finance processes. Source