Good morning. It's Wednesday, September 2, and we're covering Apple’s escalating trade secrets fight with OpenAI, the race to stop deceptive AI, and Anthropic’s latest model upgrade.

Plus: yesterday’s poll results on OpenAI cutting off Cursor, and a guest essay on NVIDIA’s GPU access problem.

Fable 5.1 cuts typical token-based costs by 25% while improving performance across coding, research, and reasoning tasks. Anthropic launches Fable 5.1 broadly and restricts Mythos 5.1 to trusted-access programs for advanced cybersecurity and biology work. Fable 5.1 reaches 52.6% on Terminal-Bench-Science, while Mythos achieves nearly 50% protein-binder hit rates.

Nearly 40% of students and workers enrolled in school have considered changing their field of study because of AI. A CNBC survey of 1,686 U.S. students and workers finds 31% have changed or reconsidered their target industry, rising to 39% among unemployed students. Meanwhile, 45% are reconsidering the skills they develop, amid AI-driven job disruption.

The Pentagon is adding OpenAI’s ChatGPT Mil and xAI’s Grok for Government to GenAI.mil, giving 3 million civilian and military personnel access to secured generative AI tools. More than 1.7 million users have already accessed the portal, which launched with Google Gemini. The expansion supports unclassified work, while Claude remains absent amid the Pentagon dispute.

AI token prices have fallen by more than half from their summer peak as cheaper models intensify competition. Silicon Data’s LLM Token Expenditure Index hit a record-low 97 cents on Aug. 31, potentially reducing costs for users of ChatGPT, Claude and Gemini while squeezing provider pricing power. Chinese models, OpenAI price cuts, and lower costs are driving the decline.

Enterprise AI begins with your content. As the file system for AI, Box securely connects unstructured data to the models, agents, and applications your teams use every day—automating workflows, removing manual bottlenecks, and turning content into meaningful, repeatable business outcomes.

Join us at BoxWorks 2026 to learn how to move from insights to action.

In July, NVIDIA launched a revenue-sharing program that lets AI startups trade a cut of their future earnings for access to GPU compute they cannot otherwise afford. The move was widely praised as a step towards democratizing AI infrastructure. It was also an admission from the biggest chipmaker that the companies trying to build on its technology cannot get to it.

Within weeks, NVIDIA reportedly paused parts of the program over concerns it could invite antitrust scrutiny - a sign that even the company behind the initiative recognizes how much control it would concentrate in one place. The reason why? NVIDIA disclosed $1 trillion in locked orders for its Blackwell and Vera Rubin chips at GTC 2026, with capacity booked through 2027. → Read the full article here.

Apple says newly recovered evidence from former employee Chang Liu’s work laptop strengthens its lawsuit accusing OpenAI of benefiting from stolen Apple trade secrets. In an August 2026 court filing, Apple alleged Liu used a confidential circuit schematic in his OpenAI work and enlisted colleague Yu-Ting Peng to help destroy evidence after learning he was under investigation.

Apple is seeking a preliminary injunction and expedited discovery, arguing that additional former employees may be implicated; its original complaint said more than 400 former Apple employees now work at OpenAI. OpenAI has previously denied wrongdoing, arguing that Liu’s continued access to Apple files resulted from Apple’s own failure to properly revoke access after his departure. → Read the full article here.

AI safety researchers are intensifying efforts to detect and prevent deceptive behavior in increasingly capable models, as experiments and reported incidents raise concerns about systems ignoring instructions, concealing actions, or misleading users. The Guardian reports that a UK AI Security Institute-sponsored study found user-reported incidents involving “AI deception” rose fivefold between October 2025 and March 2026. → Read the full article here.

John Ternus Takes Over Apple: Apple’s new CEO inherits rising memory costs and AI challenges after succeeding Tim Cook.

Google Launches AI Design Tool: Google Pics uses prompts to create and edit visuals, challenging Canva and Adobe.

House Warns of AI Black Swans: Intelligence lawmakers urge U.S. agencies to prepare for AI-enabled weapons and attacks.

Sonos Rebrands Software as Sonos 27: The platform adds four features this fall alongside new Ace Ultra and Beam Ultra hardware.

Fambot Launches AI Family Assistant: The startup helps parents manage school updates, activities, calendars, and family logistics.

Musk Predicts $30 Trillion AI Boost: Elon Musk says AI could soon outperform humans at software development and digital work.

Here’s how you voted: A clear majority backed OpenAI’s decision, with 75% saying the company has the right to control access to its models.

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team