Stripe's reported $7B+ OpenRouter deal

TechCrunch, citing Bloomberg, reports that Stripe has finalized a deal to buy OpenRouter for more than $7 billion. OpenRouter gives customers a single access point to more than 400 models; the reported price values model selection and billing as an infrastructure business.

Token brokers resell unused model credits

Matt Lenhard · Vectoral

Matt Lenhard documents brokers buying unused startup AI credits and reselling access through proxies or pooled keys, sometimes at steep discounts. His rough estimate puts tens of millions of credits on offer.

Nvidia scales back a possible $250 billion OpenAI guarantee

Reuters

Reuters, citing The Wall Street Journal, reports that Nvidia reduced the amount of OpenAI infrastructure financing it may guarantee. The change narrows Nvidia's direct exposure to one customer's data-center buildout.

Anthropic reportedly exceeds $11.5 billion in quarterly revenue

CNBC

CNBC reports that Anthropic generated more than $11.5 billion in second-quarter revenue. That demand-side figure provides a useful comparison for the financing commitments accumulating around AI infrastructure.

A PJM modeling error reportedly cost ratepayers $12 billion

SemiAnalysis

SemiAnalysis ties a planning-model error in PJM to $12 billion in U.S. ratepayer costs. The analysis offers a direct example of how capacity forecasts can move infrastructure bills.

Qwen 3.8 27B performs well locally and spends too many tokens

Simon Willison

Simon Willison's local review finds the 27 billion parameter Qwen 3.8 model capable but prone to overthinking. For agent workloads, extra reasoning tokens turn a quality gain into a latency and cost question.

A local user argues that Qwen's longer reasoning is the price of capability

LocalLLaMA

A LocalLLaMA counterargument treats the model's verbosity as the current cost of its capability rather than a defect. The disagreement gives local operators a concrete behavior to test against their own workloads.

Daring Fireball argues that Claude's watermark alters the writing

John Gruber · Daring Fireball

John Gruber argues that inserting an invisible watermark changes the writing itself. His critique tests whether text provenance can be added without changing the product users asked for.

A same-day post claims a Claude watermark bypass

Reddit · OpenAI community

A Reddit image claims a way to remove Claude's watermark without rephrasing. The bypass is unverified and should be treated as an attack sketch, not a confirmed result.

Luna Max posts 67.2% on DeepSWE at $0.61 per task

Vaibhav Srivastav

Vaibhav Srivastav reports that the model scored 67.2% on DeepSWE v1.1. He puts the cost at $0.61 per task, adding a new comparison point for long-horizon coding agents.

agent-browser pins one tab per agent across restarts

Chris Tate

Chris Tate added --pin-tab to agent-browser so multiple agents can keep separate browser tabs across commands and restarts. Persistent tab ownership makes browser state explicit in multi-agent runs.

A litigant put prompt injections into court filings

Ars Technica

A man who suspected a court was using AI inserted adversarial prompts into filings in an attempt to influence the system. The case turns prompt injection into a procedural problem for courts handling machine-read documents.

ChatGPT browser use prepares an immigration package from personal records

Atty Eleti

Atty Eleti demonstrates ChatGPT browser use preparing an immigration package from seven years of tax, banking, and immigration records. Sensitive personal records make access scope part of the product behavior.