+ OpenAI's Books Went Public, And The Numbers Are Wild In Both Directions
On August 17, reporting on OpenAI's stock market paperwork gave the public its first real look at the company's finances. The headline numbers: roughly $2 billion in revenue every month, about $25 billion a year at the current pace, against a projected loss of around $14 billion for 2026. OpenAI is targeting a stock market debut at a value above $1 trillion, with the biggest banks on Wall Street running the deal, though reports this week suggested the listing could slip from September into 2027.
The contrast with its chief rival is hard to miss. Days earlier, Anthropic reported its first profitable quarter, with revenue that has grown more than 14 times in a year. Two companies selling similar products are now telling two very different financial stories, and both are about to let the public invest.
What it all means: Until now, judging these companies meant trusting their press releases. Stock market rules change that, because filings have to hold up under legal scrutiny. Anyone whose business runs on these tools can now see what the provider actually earns and loses, and one of the two biggest names in AI is losing $14 billion a year while asking to be valued above a trillion dollars.
+ Musk's Cursor Launched A GitHub Rival On The Day GitHub Crashed
On August 18, Cursor, the AI coding company SpaceX finished buying for $60 billion just three days earlier, launched Origin, a new home for storing and collaborating on software code. That puts it in direct competition with Microsoft-owned GitHub, the site where most of the world's software teams keep their work. The timing was brutal: the same day, GitHub suffered a worldwide outage that degraded the site for more than six hours, with nearly one in five requests failing.
The outage was not a one-off. One analysis counts 257 GitHub outages over the past year, and GitHub's own technology chief admitted the platform "wasn't built for the scale it's now being asked to handle." Much of that new scale comes from AI coding agents hammering the site with automated work, which is exactly the traffic Origin says it was built for.
What it all means: We covered SpaceX's $60 billion offer for Cursor back in April, and this is what the money was for. Microsoft has owned the ground software gets built on for years, and a competitor with Musk's resources just moved in on the worst possible day for the incumbent. When a service everyone depends on stumbles this often, someone eventually builds the alternative.
+ Claude Designed New Drug-Style Molecules, And A Real Lab Proved They Work
Proteins are the tiny machines that do almost everything in your body, and most modern medicines work by grabbing onto one of them and changing what it does, the way a drug blocks a virus or shuts down a growth gone wrong. On August 19, Anthropic published results showing Claude invented new proteins from scratch, each one shaped to grab a specific disease-related target. Claude ran the project largely on its own for 48 hours. It planned the designs, operated specialist scientific software, judged its own results, and kept refining without a human steering.
Then came the real test. A laboratory physically built 1,320 of Claude's designs and checked whether each one actually stuck to its target. 354 of them did, and Claude succeeded against 14 of the 15 targets it was given. That success rate, roughly 23 to 27 percent, is about double the 10 to 15 percent human-led teams typically manage. On one target, human teams in a recent open competition succeeded 3.7 percent of the time. Claude hit 40 percent, and its single best design beat the competition's winner.
What it all means: Finding a molecule that grabs the right target is the slow, expensive first step of nearly every new drug, and it normally takes months of trial and error. Claude did it in two days at double the usual success rate, with a physical lab confirming the results. Anthropic is keeping this ability restricted in its most powerful model while it builds safeguards, because a tool that can design helpful proteins could be pointed at harmful ones too.
