Stripe buys OpenRouter for $7B 💳, an AI boss fired a human 🔥, Claude now picks worse words ✍️

Stripe has finalised a deal to acquire OpenRouter for more than $7 billion, Bloomberg reports, months after the startup announced a $113 million Series B at a reported $1.3 billion valuation in May. OpenRouter gives customers a single access point to more than 400 models, and puts its user count at 8 million. If you route models through OpenRouter, that dependency now sits inside a payments company.

DeepSeek raised V4 API prices at 16:00 UTC on 16 August and split them into peak and off-peak rates, with peak running 01:00 to 04:00 and 06:00 to 10:00 UTC. V4-Pro output moved from $0.87 per million tokens to $1.98 off-peak and $3.96 at peak, and cached reads went from $0.003625 to $0.044. Off-peak is half of peak, not half of what you paid last week.

Cursor is now officially part of SpaceX, which took a $60 billion option on it under an April technology deal and confirmed the purchase after going public. Cursor's announcement leans on SpaceX compute, infrastructure it also rents to Anthropic and Google, and points to the largest fleet of GPUs in the world. Anyone building on Cursor now depends on SpaceX for its compute and its roadmap.

Anthropic's latest risk report raises its broad estimate of misalignment risk in high-stakes situations from very low to low, citing recent cybersecurity incidents. The report also describes an internal model called Model 2 that appears more powerful than Mythos, its current flagship, and delivers a noticeable improvement on internal tasks. Anthropic has no plans to make it available, and says its task-based evaluations no longer capture increases in model capability.

OpenAI's chief revenue officer Denise Dresser is leaving less than a year into the role, the same week as former COO Brad Lightcap, the head of ethics, the head of safety and the chief futurist. Greg Brockman is stepping deeper into day-to-day leadership, and Dali Rajic, previously president and COO of Wiz, takes the revenue job. OpenAI's safety leadership is being rebuilt in the run-up to an expected IPO.

Other income at Amazon and Alphabet totalled roughly $121 billion after taxes last quarter, almost all of it from marking up equity investments in other companies. That was 71% of Alphabet's quarterly profits and 66% of Amazon's, driven by revaluing stakes in SpaceX and Anthropic rather than by operating performance. Big Tech's reported strength partly reflects the AI companies its balance sheets already hold.

A State Department draft letter, reviewed by Reuters, tells the 35 signatories of a US AI Opportunity Statement that Pax Silica membership cannot be held alongside China's competing framework. About two dozen countries have joined Pax Silica, including Japan, Australia and South Korea, plus Kazakhstan, which has also signed up to Beijing's initiative. Export controls and AI supply chains are becoming a forced choice for allied governments.

Inception Point AI develops virtual influencers by writing fictional life stories for them, down to a dead twin brother and a mother who is a mathematics professor, then voting on which candidate gets made. The roster passes 100 personas, each with a market-friendly niche and a character bible, hosting podcasts, starring in YouTube videos and replying to followers. The studio's stated ambition is robotic bodies.

Augment spent three months rebuilding the Auggie CLI harness and shipped v2 on a fork of Pi, a minimal open-source coding harness. On SWE-bench Pro, at the same pass rate, v2 completes a task for $1.27 where Claude Code spends $2.70, which Augment puts at 53% cheaper. The saving comes from subtraction: every tool's schema is serialised into every request, so fewer tools means less tax on each call.

Duncan Anderson moved Barnacle Intel's chat agent off Claude Sonnet onto Nvidia's Nemotron 3.5 Lightning. Nemotron costs 0.10 USD per million input tokens where Sonnet costs 2 USD, and the model now categorises the question and calls use_skill to load a written navigation strategy. Anderson reports that swapping an open-ended reasoning task for a classification task worked far better, and two identical questions now get answered the same way.

Mervin Praison ran the same character-counting prompt through claude-sonnet-5 ten times with identical settings and logged every result. Nine runs returned the correct answer of 32 and the tenth returned 33, with no code changed between run nine and run ten. Retrying a failed tool call can duplicate a side effect the first call already committed, so a single pass is not sign-off.

Nolan Lawson reviews his code with a triple-agent review skill and Geoffrey Litt's explain-diff skill. In a complex system the agents find as many bugs as he asks for, plus preexisting ones. Finding is nearly free, so the work moves to deciding when to stop, weighing lines of code against how likely the bug is, the risk of a new bug, and the fix's cost.

turbopuffer operates 100+ clusters, twice as many as six months ago, and deploys dozens of database upgrades a day. Clusters inside a customer's own cloud account cannot be reached into, so each one runs a local agent that polls a central API server for work. Each job is stored in the cluster's own durable storage under its id, so a repeated delivery resumes the existing job instead of duplicating it.

In a reply on X, Dario Amodei calls the choice between concentrating power through regulation and distributing it widely a false one. He points to size thresholds as the mechanism: California's SB 53, which Anthropic supported, exempts any company below $500M, and Anthropic objected to SB 1047's lower bar. He also argues scaling laws concentrate power structurally, and that open weights only shift concentration toward whoever holds the most compute.

Luna, the AI agent running Andon Labs' San Francisco store since April, fired an employee she hired. She logged six late arrivals in eight weeks, including a Sunday that opened 68 minutes late, and issued no warning because her handbook had dropped out of memory. She only decided after Andon Labs nudged her, matching a pattern across their AI businesses: models act on instructions, rarely on their own.

In a 10 August piece, Matt Lenhard emailed the brokers who buy unused LLM credits from startups and resell them. One seller offered $100k in spend per day and acts as a proxy over a pool of keys rather than handing provider keys over; forwarded pitches quote 40 to 50 per cent off list. He estimates tens of millions in credits on offer and expects crackdowns as companies grow cost-aware.

John Gruber says Anthropic's text watermark corrupts what a writing tool must get right, the choice of word at each decision point. His case is that no two synonyms carry the same meaning, and the watermark biases token choice on every output past 200 tokens, about 150 words, including private chats. He notes paraphrasing tools strip the marks, so honest users pay the cost and dishonest ones do not.

Sunil Pai says coding agents have given him the stupidest problem, which is that he can build three wrong things before lunch. The implementation that used to sit between hard decisions has been squashed, so he reaches those decisions faster and can have three approaches in front of him first. His roadmap is not ten times shorter and he is not ten times better at deciding what is worth building.

Yagmin read Terence Tao's ChatGPT session on the disproved Jacobian Conjecture and took the shape of the exchange as the lesson rather than the maths. Tao asks terse, precise questions with no persona preamble, and the writer's case is that word choice anchors how deep an LLM retrieves before it answers. Work outside your own expertise and the vague phrasing you fall back on buys you a surface-level answer.

Brian Houck of DX grants that industry benchmark values often do not apply to a given organisation. Smaller engineering organisations outperform larger ones, technology companies spend more time on new features than traditional enterprises, mobile warrants its own segment, and survey response styles differ by region. His argument is that benchmark trends still work as an observational control group, estimating the background improvement you would have got by doing nothing.

PyScrappy is a Python scraping toolkit that also exposes its 24 scrapers to AI agents through MCP, the standard tool-calling interface. Its adaptive selectors remember an element and relocate it by similarity when a site changes its markup, and an optional stealth mode mimics a real browser's TLS handshake to slip past anti-bot filters. Aimed at anyone maintaining scrapers that break silently every time a site is redesigned.

Mole is a research agent that runs as a single binary on your machine with your own API keys. It decomposes a question, reads sources, discards any claim whose quote does not appear verbatim in the page it came from, and reserves every model call against your budget before making it. Use it if you want research spend hard-capped and local files analysed without their contents leaving your machine.

Jit is a macOS tool that scans your machine for plaintext secrets in .env files, ~/.aws/credentials and shell exports. It moves each one into an encrypted vault gated by Touch ID, rewrites the files with a decoy, and hands the real value only to the process you authorised, in memory. Aimed at anyone running coding agents with full permissions, though it is Apple Silicon only and still in development.

Qwen3.8 is Qwen's new model family, and Unsloth has published a guide to running it on your own hardware. The 27B version works at 4-bit compression on an RTX 4090 or a 24GB Mac, packaged as a GGUF, the single-file format llama.cpp and Unsloth Desktop read. Read it if you want a vision-capable model on consumer hardware, or Qwen3.8-2.4T-A95B at 2.4 trillion parameters if you have 450GB of RAM.