AI Model Gateway Buyout | AI Field Notes #87

AnalysisA frontier AI lab turned an operating profit for the first time, denting the theory that compute costs make these companies structurally unprofitable. Anthropic told investors preliminary second-quarter revenue topped $11.5 billion, roughly 14 times the $787 million it booked a year earlier and more than double the first quarter, with positive adjusted operating income. Run-rate revenue has passed $30 billion, up from about $9 billion at the end of 2025. The word doing quiet work here is adjusted, which sets aside stock compensation and the training runs that built the models earning today's money.

AnalysisThe neutral switchboard developers use to dodge model lock-in now belongs to a payments company. Stripe agreed to buy OpenRouter, the gateway that routes a single API call to more than 400 models from OpenAI, Anthropic, and others, for over $7 billion. That is more than five times the $1.3 billion valuation OpenRouter carried in a $113 million round barely three months ago. Founded in 2023, it reached 8 million users by staying vendor-neutral. Stripe is betting the toll booth between apps and models becomes one of the most valuable positions in the AI economy.

AnalysisAn AI assistant that quietly records everything you do on your machine arrived as a convenience feature. OpenAI's new Computer History, an opt-in setting in the ChatGPT Mac app, logs your clicks, keystrokes, keyboard shortcuts, and app switches, then turns them into a searchable timeline that ChatGPT and Codex can draw on. It skips screenshots, audio, and incognito tabs, and stays off until you enable it. Reporters noted the captured activity is stored in plain text. The bargain is familiar from every memory feature: the more the assistant watches, the more useful it gets, and the more it knows.

AnalysisRunning a model on data you are never allowed to read has been a cryptographer's party trick, too slow and too fiddly for real products. Google open-sourced HEIR on August 17, a compiler meant to make it ordinary. Short for Homomorphic Encryption Intermediate Representation, it lets a developer write a program in Python, mark which values are secret, and compiles the rest into computation that runs on encrypted inputs without decrypting them. Homomorphic encryption means the server does the math while the data stays sealed. The pitch is that a small team can now attempt encrypted inference without hiring a cryptography group.

AnalysisAn image model that developers wired into production went dark on schedule, and the replacement costs more and needs a code rewrite. Google shut down three Imagen 4 endpoints on August 17, months after flagging the date, and now points callers to Gemini 3.1 Flash Image. The migration is not a drop-in swap: the old generate_images call is gone, replaced by generate_content, and a thousand images run about $0.067 against the $0.02 to $0.06 Imagen 4 charged. Every managed API carries this clause in the fine print, and it comes due on the vendor's calendar, not yours.

AnalysisThe lab that made its name as the cheap frontier option just built a rush-hour meter. Effective August 17, DeepSeek raised API prices across its V4 line and split them into peak and off-peak rates, a day after launching V4 Pro, which scores 53 on the Artificial Analysis Intelligence Index against 40 for the Flash tier. V4 Pro output climbed to $3.96 per million tokens at peak from a flat $0.87, and the full range runs 50 to 1,100 percent higher depending on model and hour. Capability went up while the flat, predictable discount that drove adoption went away, and the second change is the one that lands on next month's invoice.