Anthropic now charges for safety-blocked Claude requests
Three high-confidence classifiers now bill blocked calls at input rates; cyber and general-harm stay free. Plus a code-review IDE and ChatGPT voice plugins.
Copy markdown
You now pay for blocks Claude never answers
As of Sept 24, Anthropic bills pre-output refusals in three classifier categories — bio, frontier_llm and reasoning_extraction — at the refusing model's input rate with zero output tokens (a 412-token refusal runs about $0.00412 on Fable 5.1). Cybersecurity and general-harm blocks stay free, and Anthropic says these classifiers sit under a 0.1% false-positive rate with 99.7% of tested accounts untouched — but automated eval and red-team suites that trip them will now land on your bill.
An open-source IDE built to vet agent code
Whiteboard (YC W26) shipped today: an MIT-licensed, CodeOSS-based desktop IDE where you and coding agents share one workspace, with a Rust semantic-diff viewer that summarizes big functions as pseudocode and a decision log recording what the agent changed and why. It hooks into Claude Code and Greptile, runs free and self-hosted, and targets the 'cognitive debt' of merging AI output you never really read.
Your ChatGPT plugins now work hands-free
OpenAI put plugins into Live Voice on web, iOS and Android and brought voice to ChatGPT Work, so a connected app can be called mid-conversation and voice can draft docs or drive a browser. Voice-in-Work needs Plus or higher; Live Voice minutes scale from about 3 hours on Go and Plus up to unlimited on Pro $200. If you ship a ChatGPT app, it just gained a spoken entry point.