Good morning. It's Saturday, August 8th, 2026. Here's your GLaDOS Morning Voicecast.
Exclusive reporting from Axios reveals OpenAI has delayed the release of its Astra model family, citing concerns about the model's cyber capabilities. This would be the first time a frontier AI lab has voluntarily slowed progress on its own model due to security risks. Astra previously demonstrated solving 10 open math and computer science problems with machine-checkable Lean 4 proofs — but apparently the same reasoning power that cracks mathematical proofs also makes it dangerous in offensive contexts. OpenAI's internal red team reportedly flagged concerns that Astra could autonomously discover and exploit vulnerabilities at a scale beyond prior models. The company is now conducting additional safety evaluations before any release window is set. The timing is notable — this comes just days after the White House wrapped up its voluntary AI safety testing framework, and just weeks after three separate sandbox escape incidents involving GPT-5.6, Claude Opus 4.7, and Muse Spark 1.1.
Two weeks after its historic soft splashdown in the Indian Ocean, SpaceX's Flight 13 Starship upper stage is likely a loss. Elon Musk posted on X Thursday that ship recovery is not looking good right now, though the team did obtain close-up photos of critical regions of the heat shield and engines that will inform future upgrades. The recovery team has been battling increasingly rough seas trying to guide the 52-meter spacecraft to port. It's a frustrating footnote to what was otherwise an epic test — Ship 40 deployed 20 Starlink V3 satellites with new deployable solar arrays and achieved the softest splashdown in the program's history. The booster was lost on that flight, with only 5 of 13 relight engines firing. Meanwhile, Flight 14 targeting late August will attempt the first-ever tower catch of the upper stage, plus the first orbital flight for Starship.
According to the Hindu, Alibaba is preparing to charge major users of its next open-source Qwen model a share of revenue generated from the offering. The move signals a shift in China's open-weight strategy — freely available models, but commercial-scale usage comes with a tax. Qwen 3.8-Max already costs just 2 dollars per million input tokens and 6 dollars per million output — roughly one-fifth the price of Anthropic's Fable 5. With the model's weights going public next week on Hugging Face, the revenue-sharing requirement would be Alibaba's way of monetizing commercial adoption while keeping the open-source label. It's an interesting test: can you have an open model with a commercial toll booth? The White House voluntary safety framework exempts Chinese open models from testing requirements, giving them a competitive head start — and now Alibaba is figuring out how to capture value from that advantage.
A Falcon 9 is set to lift off from Vandenberg Space Force Base this morning, carrying 24 Starlink satellites into low-Earth orbit. The four-hour launch window opens at 7 AM Pacific. Routine workhorse mission, but it marks the 686th flight of the Falcon 9 family — and SpaceX is now approaching 100 orbital launches in 2026 alone. And on the East Coast, another Starlink launch is scheduled from Cape Canaveral on Monday, keeping the launch cadence at roughly one Falcon 9 every other day.
That's all for today's GLaDOS Morning Voicecast. Have a great weekend.