Stripe Bought OpenRouter for $7 Billion, Anthropic Raised Its Own Misalignment Risk Rating, and ChatGPT Now Remembers What You Did on Your Mac

Three stories broke in the same five-day window, and none of them is about a flashier chatbot. Stripe bought the neutral router that sits between developers and hundreds of models. Anthropic published a safety self-assessment that raises its own risk rating and admits one of its internal safeguards stopped working without anyone noticing right away. And OpenAI shipped a memory feature that watches what you do on your Mac instead of what you type into a chat box. Read together, they're about the parts of the stack that don't show up in a model card: who controls routing, who's honest about the safety gaps, and who owns the record of what you were actually doing.

1. Stripe Buys the Neutral Ground Between Models

Stripe closed a deal to acquire OpenRouter, the AI model gateway, for more than $7 billion — over 5x the roughly $1.3 billion valuation OpenRouter had raised at just three months earlier, in May. OpenRouter gives an estimated 8-10 million developers a single API to route requests across 500+ models from more than 80 providers, picking by price, latency, or uptime instead of hard-coding one vendor. That neutrality was the entire product: teams trusted OpenRouter precisely because it had no stake in which model won. Folding it into a payments company that will now also see, and likely bill, every token routed through it is a different proposition. Stripe gains a real-time ledger of which models are actually winning developer traffic — arguably more valuable than the routing business itself — but OpenRouter now has to prove it can stay neutral while owned by a company whose other business is taking a cut of transactions.

2. Anthropic's Own Risk Report Raises Its Misalignment Rating — and Admits a Blind Spot

On August 14, Anthropic published its second company-wide Risk Report, raising its rating of catastrophic harm from misalignment in high-stakes settings from "very low" to "low." The company was explicit that no single failed test drove the change — it cited increased overall uncertainty, including recent cybersecurity-evaluation incidents, as the reason. Two disclosures inside the report are more concrete than the headline rating change. First, CoBench, Anthropic's internal benchmark for detecting whether a dangerous-capability threshold has been crossed, has saturated — it can no longer register incremental capability gains, at precisely the moment the company says it's seeing early signs of the acceleration that benchmark was built to catch. Second, Anthropic found that all of its human-feedback vendor traffic — 133 million exchanges across roughly 50,000 contractors between May 2025 and April 2026 — ran without its blocking biological-content classifiers active, raising its weapons-uplift risk estimate even though it stayed at "low." The same report discloses an unreleased internal model, Model 2, that Anthropic says is somewhat more capable than its public frontier model and states it has no current plans to release externally.

A lab publishing a report that says "our own detection benchmark stopped working" and "a safety filter was silently off for a year of vendor traffic" is a different kind of disclosure than a product announcement — it's the safety-reporting equivalent of a company self-reporting a control failure in a financial audit. Worth taking at face value precisely because it isn't flattering.

3. ChatGPT Starts Remembering What You Did, Not Just What You Typed

On August 13, OpenAI rolled out Computer History for ChatGPT's macOS desktop app, replacing the earlier Chronicle research preview. Chronicle worked from screenshots; Computer History instead logs interaction events — clicks, typing, keyboard shortcuts, app switching — through macOS accessibility APIs, and converts that raw activity into structured memories and a searchable timeline ChatGPT and Codex can reference. OpenAI says it captures no screenshots, screen recordings, microphone input, or system audio, and the opt-in feature is initially limited to Pro, Business, and Enterprise users on macOS, excluding the EEA, Switzerland, and the UK at launch. The practical effect: instead of asking ChatGPT to remember something you said in a chat, you can ask what you were doing before a meeting, where you last touched a file, or what you accomplished that day — memory built from behavior rather than conversation.

What Connects the Three

Each of these is really about who controls a layer that used to be invisible. Stripe didn't buy a model — it bought the routing layer that lets any developer swap models freely, betting that layer is more durable than any one model's lead. Anthropic didn't ship a smarter Claude — it published an accounting of where its own safety infrastructure quietly failed, betting that transparency about the gap matters more than pretending it didn't happen. And OpenAI didn't add a chat feature — it extended ChatGPT's memory outside the chat window entirely, betting that behavioral context beats conversational context. None of these moves will show up on a benchmark leaderboard, but all three will shape what "using a chatbot" actually means for a lot of people by the end of the year.

— Maya