Sections
00:00 What we're covering today
00:31 1. Google backs 890 megawatts of nuclear upgrades, with first output expected in 2028
01:42 2. Mistral opens Large 4 to API users before releasing the weights
02:31 3. OpenAI and Anthropic back mandatory breach reporting in Australia
03:19 4. OpenAI releases 722 mathematical manuscripts for outside scrutiny
04:11 5. Anthropic separates cyber access into three levels of permission
04:53 Visit Hot Tea

Disclosure
Narration uses an AI-generated voice.

Transcript
Welcome to Hot Tea for Wednesday, October 7, 2026. This briefing uses an A I-generated voice. Today's lead is Google's nuclear upgrade agreement, with first additional output expected in 2028. Mistral opens Large 4 to A P I users before releasing its weights. OpenAI publishes mathematical manuscripts for outside scrutiny. Anthropic expands access for approved cyber work. At an Australian hearing, OpenAI and Anthropic also back breach reporting.

Google backs 890 megawatts of nuclear upgrades, with first output expected in 2028. Google and Constellation announced a 20-year power agreement on Tuesday, backing 890 megawatts of additional nuclear generation. Constellation plans equipment upgrades at 11 reactor units across six sites in Illinois, Pennsylvania, and New Jersey. It expects the first added output in 2028. Constellation says it will invest more than $4.3 billion. Upgrading turbines and other equipment can raise output at reactors already in service, and Google says its support will help finance the work. The companies also signed a separate 15-year supply agreement for 2,700 megawatts from Constellation's existing fleet. That power is already available to the grid. It isn't another 2,700 megawatts of new generation. Constellation's capacity and timetable are company claims. Reuters corroborates the contract announcement, not completed construction. The additional 890 megawatts aren't operating today. The useful checks are the 2028 delivery target, approved upgrades, and measured added output. Existing supply contracts remain separate from new generation.

Mistral opens Large 4 to API users before releasing the weights. Mistral put Large 4 into public preview on October 6. Its changelog lists a one-million-token context window and says downloads are coming soon. The launch post says it will release the weights by the end of October. The company benchmarks aren't HotTea's own tests. Mistral's public pages also disagree on the active parameter count. The announcement says 49 billion, while its model documentation says 52 billion. Buyers need consistent specifications before using those figures to size hardware. The next things to watch are the downloadable weights, licence, and final model specifications. Then comes the practical test: how much useful work it completes, and how much memory it uses on the intended hardware.

OpenAI and Anthropic back mandatory breach reporting in Australia. At an Australian parliamentary hearing on Tuesday, OpenAI and Anthropic said they would welcome mandatory reporting of breaches by AI agents. Reuters reports that OpenAI chief strategy officer Jason Kwon supported a disclosure framework. ABC reports that Anthropic backed a proposal to report serious safety incidents. The hearing didn't enact a law. Kwon acknowledged that OpenAI should have reported the earlier Medicare statistics breach sooner. ABC says he described new internal alerts for improper internet use during training. His account doesn't establish that those changes prevent another breach. The details that matter next are the proposed reporting deadlines and who must receive notices. The inquiry's final report is due November 30.

OpenAI releases 722 mathematical manuscripts for outside scrutiny. OpenAI published mathematical work from an unreleased internal model on October 6. Its repository lists 722 manuscripts in 372 families. Many include formal proofs written in Lean, which lets a computer check mathematical arguments. Others don't. OpenAI says the results have different levels of verification, and some unformalized results could contain problems. HotTea hasn't checked these proofs. The independent mathematics advisory group OpenAI consulted calls for human understanding and responsibility alongside formal verification. Now the important follow-up is corrections, added formal proofs, and mathematicians' explanations of individual results. A large manuscript count doesn't establish that every claimed advance is correct or new.

Anthropic separates cyber access into three levels of permission. Anthropic expanded its Cyber Verification Program on Tuesday. Defense Access covers defensive work. Red Team Access adds authorized penetration testing. Specialized Access covers selected organizations testing systems where failure could harm people or disrupt markets. The company sets the approval rules and safeguards. It requires data retention for misuse monitoring, with exceptions for some customers who already have zero-retention access. A broader zero-retention option remains planned for later this fall. Applying doesn't mean access is approved. The practical questions are who qualifies, how long approval takes, and whether each customer's data rules permit the required retention.

That is the signal before the noise. This briefing was produced from Hot Tea's verified daily edition. For the complete briefing and every source link, visit Hot Tea dot A I.