Top story
OpenAI's September revenue run rate was near $50 billion
The Financial Times reviewed investor documents showing annualized revenue approaching $50 billion at September's end; CNBC independently confirmed the figure with a person familiar with OpenAI's investor presentation. Both accounts say the larger number arose from an attempt to compare OpenAI's revenue with Anthropic's using different accounting treatments.
Why it mattersInvestors are weighing OpenAI's growth against enormous compute commitments, and the clarification coincided with declines in AI-linked stocks.
Also
Anthropic launches a critical-infrastructure cyber program
The company says it is pairing Claude, on-site engineers and threat research with security firms including CrowdStrike and Dragos to help operators find and fix weaknesses.
Why it mattersOperators can work with trusted security providers to test Claude-assisted vulnerability repair on systems that are difficult to take offline.
Anthropic updates Claude's usage policy on deceptive campaigns and model abuse
The update, effective November 12, consolidates existing bans on deceptive campaigns and clarifies that Claude cannot be used to deceive voters or disrupt elections. It newly prohibits sustained, purposeless cruelty toward the models, while dropping the blanket ban on personalized vote and campaign targeting.
Why it mattersThe rules distinguish deceptive election activity from legitimate voter information and spell out limits for users and developers.
Anthropic commits $150 million over three years to the Genesis Mission
The company says the support will make Claude, Claude Code, API credits and training available to several hundred research projects across more than 15 participating federal agencies, including NASA and NIH.
Why it mattersResearchers at those agencies will have a funded route to test Claude on scientific work.
Google announces a cross-app Gemini agent for enterprises
Google's new agent is designed to carry out assigned work across connected business systems, delegate to subagents and choose between Gemini and Claude models; The Verge says access is currently a private preview.
Why it mattersEnterprise customers can test task automation under their existing app identities and controls.
Fired OpenAI safety researchers dispute misconduct allegations
Three former employees told TechCrunch that their dismissals could chill safety work; OpenAI said an investigation found a pattern of misconduct and denied retaliation for raising safety concerns.
Why it mattersThe researchers say the firings could discourage colleagues from discussing safety risks with outside evaluators.
Research worth a look
Can AI automate Epoch?
Epoch AI tested six models on 11 tasks drawn from its own work and found that frontier models handle defined assignments more reliably than open-ended research and design work.
Tools & releases
Anthropic's OSS Scanner
Eligible open-source maintainers can opt in to free periodic vulnerability reports generated by Claude models without human review; Anthropic cautions that some findings may be inaccurate.