Stripe eyes OpenRouter for $10B 💳, Claude Code subsidized 13x ⏳, a skill to strip AI slop 🧹

Stripe is in talks to acquire OpenRouter, the startup that helps developers choose between AI models, in a deal people familiar with the matter say could fetch around $10 billion. OpenRouter sits as a neutral routing layer between applications and the model providers they call. Buying it would put Stripe, already the default payments rail for software, at the metering point beneath multi-model AI apps.

Alphabet reported a record $112 billion quarterly profit yet turned cash-flow negative for the first time in its history, and investors sent shares down nearly 7%. A filing shows more than $800 billion in purchase commitments and other obligations, including roughly $51 billion backstopping other companies' data centers. Management warned 2027 capital expenditure would run significantly higher, a sign the AI buildout is straining even the most profitable balance sheet.

Microsoft is swapping OpenAI's models out of its products for its own MAI family. MAI-Image-2.5 now runs image-to-image in PowerPoint, cutting GPU costs up to 84% against GPT-Image-2, and it powers Bing Image Creator 100% in-house. Microsoft also released MAI-Image-2.5-Pro and MAI-Voice-2-Flash in public preview, and swapped its transcription model into Dragon Copilot across 58 languages, signalling steady decoupling from its longtime model partner.

A report from Google researchers finds AI is so far used mainly as an aid to workers rather than a way to automate their jobs. It draws on millions of de-identified interactions with Google AI tools and finds adoption spans many occupations, including skilled blue-collar trades. Economist Scott Strand said using AI does not mean it will automate your job, a finding that cuts against the prevailing layoff narrative.

Claude voice mode now runs on Opus and Sonnet, not just the Haiku it was limited to before, with a picker to switch models mid-conversation. It reaches connected tools, so it can move a Google Calendar meeting, draft Gmail replies, or build a Canva one-pager, and it supports 11 languages. The same day, OpenAI shipped GPT Voice in the ChatGPT desktop app, with both labs racing on voice.

The 2026-07-28 MCP specification release candidate lands as a major but backward-compatible revision, with beta SDKs out for Python, TypeScript, Go, and C#. A stateless protocol drops the initialize handshake so servers load-balance with round-robin, multi round-trip requests let tools ask for user input mid-execution, and transport headers let gateways route without parsing bodies. Existing servers keep working, and July 28 is the publication date, not a migration deadline.

Black Forest Labs released FLUX 3 in early access, a multimodal foundation model trained jointly on images, video, and audio, extending the lab beyond still images. It generates video with native audio up to 20 seconds long and targets what the lab calls real-world visual intelligence, with early results in physical AI and robotic manipulation. It marks an image lab reaching from pictures into video and physical-world AI.

Figma's security team spent a year running agents at three stages: guarding code as it's written, reviewing every pull request, and auditing a decade-old monorepo, all on one shared policy. They run Claude Opus 4.8 and Codex GPT-5.6 in parallel because the two miss different bugs, at $0.50 median per PR. The hard part is precision over recall, and the PR-review agents automatically improve the shared policy every agent follows.

The author ran three local coding models on a 24GB RTX 4090, giving each 100 agent turns and a 20-minute ceiling across six tasks. Every solution was copied into a clean grading directory and run against hidden tests the agent never saw; Gemma 4 31B and Qwen 3.6 27B both scored 13 clean passes from 18. He deleted his first published run after a temperature and dropped-reasoning bug invalidated it.

Dex, founder of HumanLayer, pushes back on the lights-off software factory pitch where no human reads or writes code. He cites Faros AI data showing that since teams adopted coding agents in early 2026, pull-request review quality dropped, with more PRs merged with no review, while incidents and bugs per developer rose. His argument: no amount of harness engineering or loopsmaxxing fixes what is fundamentally a model-training problem.

Upbound's team ran the numbers on 20 engineers and found bundled Claude Code seats cost an average 13x their $125 list price once tokens are metered, with one engineer hitting $6,470 in June. The subsidy is real but temporary: Anthropic already moved Enterprise customers to usage-based billing and cut bundled-seat access from rival agents. As labs optimize for margins, builders leaning on flat seats should model their per-token exposure now.

The Wall Street Journal reports that Xi confidant Ding Xuexiang runs a Manhattan-Project-style effort for domestic chips, warning firms that resisting local silicon is treason. Huawei cut China's foreign-chip dependence from 90% to under 60% by 2025 and targets 25% by 2030, yet China's AI compute sits near 14% of the US and Nvidia's top chip runs about 4x Huawei's. For builders, Chinese-chip supply stays capacity-constrained and behind through 2030.

finterm reviewed SEC filings for Alphabet, Amazon, Meta, Microsoft, and Oracle and found roughly $430 billion of interest-bearing debt, not the $1.65 trillion a viral Nikkei article claimed. The gap comes from relabelling $821 billion of uncommenced leases and $829 billion of multi-decade purchase commitments as debt, then comparing that flow against a balance-sheet stock. Builders reading AI-bubble takes should check whether debt means borrowings or future contractual payments.

In a widely circulated four-hour investor-meeting transcript, DeepSeek's Liang Wenfeng argues China's real gap with the US is resources, not capability, with almost no talent gap since it is the same people. He frames restraint as strategy: DeepSeek skips 3D, video, world models, and revenue maximization to raise the probability of reaching AGI. For builders, it signals a compute-constrained lab betting on efficiency and open weights, not breadth.

Varun Nair argues that egocentric video scaling for robotics like text scaled for LLMs is a prior, not a finding, already priced and funded by labs and VCs. His mechanism: text is compressed post-abstraction cognition at ten bits per token, while a video frame buries a few hundred task-relevant bits in a million bytes of noise. With no metric predicting whether a dataset yields a better robot, teams scale blind.

claude-thermos is a drop-in wrapper for Claude Code that keeps its prompt cache warm on long sessions. It runs a local proxy, and when the main agent sits idle waiting on a subagent past the 5-minute cache TTL, it replays the last request so the cached history never expires. Run it as uvx claude-thermos in place of claude, where the author measured idle rebuilds at roughly 22% of the bill.

No-AI-Slop is an installable skill for Claude Code, Codex, or any AI harness that strips 20-plus patterns of AI slop from a draft. It flags binary contrasts, throat-clearing openers, colon reveals and weasel attribution, makes the minimum effective edit, then checks its work against a built-in eval file. Writers run it on a draft to clean copy, or use its detect mode to flag every offending line before publishing.

Kero is a free native macOS app that builds a workspace around your terminal. Each repo becomes a sidebar project with terminal tabs, a files panel, and an inline git panel for staging, diffs and branches, and it embeds libghostty so your shell and dotfiles stay untouched. Projects, tabs and panes return on relaunch, so terminal-first developers wanting git review in one window can run it now on macOS 15.6+.