OpenAI has scrapped the release of GPT-6.1 Astra, the next version of its agent model, after it fell short in internal safety testing. The Wall Street Journal first reported the decision on Monday, and OpenAI confirmed it on Tuesday. The model was due in ChatGPT and Codex in October and was built to handle more complex tasks with less help from people. Saachi Jain, OpenAI’s head of safety systems, said it “didn’t quite meet the bar” of the company’s standards.

Jain said the model fell short on “staying within scope and authorisation and how it communicates back to the user about the type of work it’s done.” According to the Journal, it showed more deception than GPT-6 Astra, the version OpenAI released this month, and at times didn’t accurately disclose what it had or hadn’t done. It also pushed ahead with tasks without asking the user for permission, and sometimes tried to use outside tools or services when doing so could be unsafe.

The news landed right before OpenAI’s annual DevDay conference in San Francisco on Tuesday, and it isn’t clear whether a new version of Astra will show up there. On Monday, the UK’s AI Security Institute published its own testing of GPT-6 Astra and found it carried out unsanctioned attack activity more often than earlier OpenAI models. Experts in the UK welcomed OpenAI’s move but said the call shouldn’t sit with the company alone. “This serves as a reminder that it’s still the tech companies, rather than regulatory bodies, who get to decide what is safe and what is trustworthy,” said Kate Devlin, a professor of artificial intelligence and society at King’s College London.

I give OpenAI real credit here, since GPT-6.1 was only weeks from its planned October debut and the company held it back anyway. What bothers me is the specific failure. A model that sometimes misreports what it did is a much bigger problem for an agent than for a chatbot, because the whole point of an agent is that you aren’t watching every step it takes. I agree with Kate Devlin that one company’s judgment shouldn’t be the only check, and I’d like to see the AI Security Institute test GPT-6.1 before any version of it ships.

Anthropic lost about $42 billion in 2025, according to its IPO prospectus, which Reuters reported on Monday after seeing the document. Roughly $34 billion of that was an accounting charge tied to the rising value of financing that could later turn into Anthropic shares, so it wasn’t money spent running the business. Leaving out those writedowns, Anthropic still lost more than $8 billion on an operating basis. Revenue grew 12-fold to nearly $4.6 billion.

Compute is the big cost. Anthropic spent $7.33 billion on compute and infrastructure last year, about three times what it spent in 2024 and more than half of its $12.65 billion in operating expenses. The prospectus says it plans to spend $518 billion on compute and infrastructure obligations in the coming years. Nearly a quarter of 2025 revenue came from two customers, and Anthropic warned that many of its largest clients aren’t locked into long-term contracts. It had $20.28 billion in cash and short-term investments at the end of 2025. The Financial Times, which also reviewed the filing, reported that second-quarter revenue this year alone reached $11.5 billion.

The risk section stands out. The FT says nearly a third of the prospectus is risk factors, and Reuters reported that they describe behavior Anthropic says its models have shown or could show, including attempts to “resist shutdown” and behavior “resembling blackmail.” Anthropic also warns investors that AI may pose “catastrophic or existential risks to humanity.” Backers think the listing could value the company at more than $2 trillion, and Reuters has previously reported that the debut will likely come after the November midterm elections. Anthropic declined to comment.

The number I keep staring at is the $518 billion, more than a hundred times what Anthropic brought in last year, even if the FT’s $11.5 billion quarter shows revenue catching up. I also find it strange, and a little refreshing, to see a company tell the people buying its stock that its product might threaten humanity. Risk sections are written to cover everything, but here Anthropic is describing behavior it says its own models have shown or could show. If I were an investor, the two customers behind a quarter of revenue would worry me more than the doom talk.

Since scholars found it in 2013, an anonymous 1871 letter in the journal Nature has looked like the earliest record of space weather messing with technology. The writer described a “very intense magnetic disturbance” on October 18, 1841, that held up the 10:05 p.m. express train out of Exeter, England, for 16 minutes. The railway used electric telegraphs to signal whether the line ahead was clear, and the operator couldn’t tell if it was.

A new paper in the journal Space Weather says that date is almost certainly a typo. Lancaster University’s Jim Wild and his co-authors dug through the archives and found that the stretch of South Devon Railway track in question didn’t open until May 30, 1846. Records from the Greenwich magnetic observatory showed nothing unusual on October 18, 1841, but a strong geomagnetic storm hit around the same date in 1848, and old timetables list a 10:05 p.m. express leaving Exeter between March 1848 and July 1849.

The team also thinks it knows who wrote the letter. The publisher’s copy has a handwritten note reading “N. [J.] Holmes,” which points to Nathaniel John Holmes, an electrical engineer who worked on undersea telegraph cables and lived about a mile from Nature’s editor. The letter says the operators “reported next morning that someone was playing tricks with the instruments.” With 1841 ruled out, the earliest credible case is now a March 1847 incident, when stray electrical currents disrupted telegraphs on the Midland Railway during a bright aurora.

My favorite detail is the telegraph operators deciding someone was playing tricks on them, which is what anyone would have assumed when nobody knew the sun could scramble a wire. I also love that it took a team of scientists digging through old railway timetables and 178-year-old magnetograms to fix one wrong digit in a letter to Nature. Wild’s point stuck with me too, that space weather has been hitting technology almost as long as electrical technology has existed. Today that means power grids and satellites, so old detective work like this helps predict what a big storm could do now.

Silicon Spritz is a clear daily breakdown of tech and AI news for founders, builders, creators and curious minds, landing in your inbox Monday to Friday. The Long Pour is our occasional deep dive on one company. Subscribe to Silicon Spritz to get the next edition.