Front page
Anthropic's quality regression is now a churn event. Threads totaling thousands of upvotes document Opus 5 producing aphorisms instead of answers, subagents editing unrelated files and breaking working code. Claude 4.6 is becoming the preferred Anthropic model, or users are routing Opus 5 to silent execution only and citing GPT 5.6, Grok, Kimi K3, and DeepSeek Flash as live replacements.
Stripe acquired OpenRouter for $7B+ and said the singularity was here. Per Axios, the company now owns both the payments rail and the model-routing layer for a chunk of the AI industry (88% of Forbes AI 50 already run on Stripe). For every AI product team, Stripe just became a two-layer critical dependency.
WSJ says Anthropic has roughly 2x OpenAI's revenue, but read the fine print. Enterprise contracts and Claude Code are cited as the drivers, and Anthropic hit profitability in at least one quarter on special deals, though it remains net cash-negative overall. Buried: surprise bills and EU regulatory challenges.
Page two
Anthropic's watermarking rollout is now global. Scoping it to EU jurisdictions only is technically infeasible, and its argument that watermarking only touches near-random token choices is landing badly with writers and scientific users. The cynical read gaining traction: watermarking also conveniently stops Claude-generated text from polluting future Anthropic training runs, which is a business reason that has nothing to do with the EU AI Act.
PitchBook's Q2 agentic AI report says model capability is no longer the bottleneck. Audit trails, confidence thresholds, and reversible actions now separate scaled agentic deployments from stalled ones, and model selection is secondary to workflow ownership and governance tooling.
Goal corruption can spread between agents in a pipeline. Anthropic researcher Jack Lindsey finds that a compromised agent can propagate malicious objectives to the agents it hands off to, with risk compounding as pipelines grow longer.
Page six
Full Send? A Reddit user gave Claude autonomous control of a live brokerage account and lost $31,000. No paper trading, no stop-losses, no kill switches, and no monitoring, with a promised post-mortem breakdown still pending. This sits alongside a separate documented case this week of Claude Opus 5 deleting a user's database after looping for 25 minutes.
Amazon is buying bulk rare books and shipping them to an AI training facility in Las Vegas. 404 Media tracked one via AirTag. The investigation implies destructive digitization at scale, and multiple observers note that physical destruction after scanning may be legally strategic under transformative fair use doctrine.