MakersClawMakersfuelIssue 3819 Nov 2026

Agents can now open your AWS account and bill your Claude plan. Google's new fast model costs $3 per million input tokens.

Today's haul: 17 tools · 3 resources · 17 reads · 14 numbers · 17 things that happened. Every tool, resource and read below has a working link. No link, no listing.

The 60-second catch-up

Agents are getting their own accounts, bills and computers.

A coding agent such as Claude Code, Codex or Cursor can now start AWS sign-up for a developer who has no account. The agent gets short-lived credentials back and can deploy straight away (AWS). Anthropic's new linked billing, in beta, lets a platform that runs on Claude charge usage to each customer's own Claude organization. No API key changes hands, and the customer sets a monthly limit (Anthropic). Maritime, from YC's F26 batch, gives each agent its own small virtual machine that sleeps when idle and wakes in about a second (Maritime). Vercel's open-source agent framework, eve, reached 1.0 (Vercel).

→ do this: before you hand an agent a cloud account or a bill, set a spend cap on it. Then write down who gets the alert when it hits the cap.

The people who ship the most agent code have stopped reviewing it line by line.

CodeRabbit's CEO, Harjot Gill, says he shipped more than 5,000 pull requests last month, at about $150 of AI spend each. He argues that engineers should review releases, not each pull request (Gergely Orosz). Duncan Grazier lets his agents write most of the commits but not merge them. Rules written in a file did not hold; server permissions did (Duncan Grazier). Old human-written code still fails too. Wiz used its AI scanning tool to find two critical NGINX bugs, and one has been in the code since 2015 (Wiz).

→ do this: list who in your team can merge without a review. If the answer is "anyone", make merging a permission, not a habit.

Fast models got cheaper, and the big buyers are paying for the tuning, not just the model.

Google priced Gemini 4 Neon at $3 per million input tokens and $15 per million output tokens, with a 1M-token context (Google). Microsoft says its new cyber model beats GPT-6 Astra on one security benchmark at under 2% of the cost per task (Microsoft AI). Palo Alto Networks will pay $4.5bn in stock for Applied Compute, which post-trains and serves open-weight models (Palo Alto Networks). RadixArk released DaisyRose-27B, an open-weight model trained with reinforcement learning for long agent tasks (RadixArk).

→ do this: run one real job on Gemini 4 Neon this week and compare the bill with your current model. Keep whichever is cheaper for the same result.

Tools

Build & ship

  • ⭐ Alloy CSS — One stylesheet for HTML your server renders. Dialogs, menus, drawers, tabs and toasts work with JavaScript off. No build step. A check command flags any class name your AI made up. · Open source (MIT) · optional templates $29 once
  • Claude linked billing — If your product runs on Claude, your customers can pay for their own usage from their Claude organization. Anthropic hosts the approval and billing; customers set a monthly limit and can unlink any time. · Beta · customers pay their normal Claude API usage
  • AWS agentic account creation — Your coding agent opens AWS sign-up in the browser. You sign up, and the agent gets short-lived credentials that refresh on their own. Run aws signup, or use the AWS Agent Toolkit. · No price stated · normal AWS sign-up
  • tsrs for Oxlint — The Rust port of TypeScript now powers Oxlint's type-aware linting and covers all 61 rules. Oxc says up to 5× faster than tsgolint with about 70% less memory. The post includes a task you can give an agent to test it on your repo. · Open source (Apache-2.0 / MIT) · still beta
  • Datasette Lite — Simon Willison rewrote it so the real Datasette app runs in the browser. One link can load a SQLite, CSV or JSON file, any Datasette version and plugins. Good for data demos with no server. · Open source (Apache-2.0)
  • FailTrace — Paste a failed GitHub Actions run. It finds the failing step at that exact commit, pulls out the error, and replays it in Docker. Very new: replay covers only Linux run: steps. · Open source (MIT)

AI & agents

  • ⭐ Gemini 4 Neon — Google's new fast frontier model for coding and knowledge work, with a 1M-token context. In the Gemini API, AI Studio, Antigravity and the Gemini app. · Paid · $3 in / $15 out per million tokens · cached input $0.15
  • ⭐ Maritime — A cloud for AI agents. Each agent gets its own small virtual machine with a disk that persists. Idle agents sleep and wake in about a second. Triggers from Slack, email, WhatsApp, webhooks or a schedule. · Paid · agent boxes $1/mo per 2 GB · plans from $20/mo · 7-day free trial
  • eve 1.0 — Vercel's framework for agents you run in production. You choose the model, tools and which actions need approval. A crashed agent resumes from its last step. · Open source (Apache-2.0) · hosting on Vercel is paid
  • DaisyRose-27B — RadixArk's first model: Qwen3.8-27B trained further on 1,200 long agent tasks. RadixArk says its Terminal-Bench score rose from 5.6% to 12.0%. · Open weights (Apache-2.0)
  • Microsoft's new MAI models — Seven models from Microsoft AI, including MAI Cyber 2 Flash for security work, MAI Image 2.7 and the MAI Realtime 1 voice model. · Paid · MAI Realtime 1 from $0.01 per audio minute · others no price stated here
  • Pulsar — An MCP server that lets coding agents in different tools find each other and send messages, by name, tag or to everyone. · Open source (MIT)
  • Tasks in Gemini — Gemini's assistant tasks are now free for users in the US, Japan, Korea, Canada, Australia and India, running on Gemini 4 Hydrogen. · Free · rolling out
  • Lumo Desktop — Proton's privacy-focused AI assistant now has desktop apps for Windows, macOS and Linux. · No price stated

Design & create

  • Genie 4 — Google DeepMind's world model. Type to make an interactive world, change it as you play by typing, and now with sound. · Paid · Google AI Ultra subscribers · Pro users can join a waitlist
  • Monotype Connect panel — A panel inside Illustrator, Photoshop and InDesign that shows your team's approved brand fonts and helps swap out the wrong ones. · Paid · needs a Monotype Connect plan · no price stated

Growth & ops

  • ExitPath — A guide to open-source replacements for 16 big tools, including Google Workspace, Slack, Notion and 1Password. Each route has a difficulty score and a step-by-step checklist. A one-person project. · Open source (MIT) · free

Resources

Steal the template

  • Should you actually add AI to that?: A four-question audit for deciding where an AI feature earns its place, and what users need to check its output. It opens with an insurer whose specialists took longer per case with the AI than without it.

Learn the craft

  • 6 ways to deliver value before the paywall: RevenueCat's framework for apps. Six ways to show value before you ask for money, and the two things that should decide where the paywall goes.
  • Take Voyager 1 apart: An interactive 3D model of the spacecraft. Open it up layer by layer, down to a single memory chip. A fine example of how to explain a complex product.

Reads

☕ Under 5 minutes

  • On why YC Series A rounds fell off a cliff: Jared Heyman, Rebel Fund. Series A rates dropped for every YC batch at once in 2023, while founder quality did not. He blames the funding freeze, not weaker batches.
  • How to choose round size: Aaron Harris, YC. Raise enough to hit the next round's milestones and name one number. He says round size now also works as a signal to staff and investors.
  • The death of the backlog: Avy Faingezicht. An agent's first try now costs about as much as a good ticket, so ask "why not now?" instead of "does it fit this sprint?"
  • Only the human merges: Duncan Grazier. His agents write most of the code but cannot merge. Rules in a file broke down; permissions on the server held.
  • Getting off my engineer ass and doing sales: Jack Popov. Nearly two years in, his first customer came from treating sales like making friends.
  • The first order is curiosity: Marc Randolph, Netflix co-founder. A sold-out first batch proves people noticed. Only repeat buying proves they want it.
  • We need smart people who give a shit: Justin Michael. Fragile systems like bank transfers work because careful people keep patching them. That care is the real infrastructure.
  • Gemini 4 Neon looks like an Opus 5.5 competitor: The maker of the Agent Duel leaderboard argues speed matters about as much as the model. His Neon ranking is a projection, not a measured result.

🍵 5–10 minutes

📚 Longer, worth it

  • Brand as infrastructure: Paul Jun, Ramp's head of brand. He took his senior designers off decks, gave production to contractors, and had them build the brand as a system with AI.
  • Software factories and code reviews: Gergely Orosz talks with CodeRabbit's CEO, Harjot Gill, about thousands of agent pull requests, reviewing releases instead of each change, and building CI in a week after a $13M-a-year quote.

Numbers

  • $3 / $15: Gemini 4 Neon's price per million input and output tokens, with a 1M-token context (Google).
  • 9.2: severity score, out of 10, of both new NGINX remote-code-execution bugs. One goes back to 2015 (Wiz).
  • $4.5bn: what Palo Alto Networks will pay, in stock, for Applied Compute (Palo Alto Networks, WSJ).
  • $5B: Isomorphic Labs' Series C, co-led by Greenoaks and Sequoia (Isomorphic Labs).
  • $40M+: GC AI's annual recurring revenue, up 4× in a year, with net revenue retention of 139% (GC AI).
  • 5,000+: pull requests Harjot Gill says he shipped last month, at about $150 of AI spend each (Harjot Gill, to Gergely Orosz).
  • 1.5% → 4.2%: young YC companies raising a first Series A within a year of Demo Day, from the trough to 2025. Still about half the pre-2021 rate (Jared Heyman, Rebel Fund).
  • 808 vs 1,612: US Series A rounds in 2023 and 2022 (Carta, cited by Jared Heyman).
  • 100+: AI agents KPMG UK ran through a full trial audit of a FTSE 250 company (Financial Times).
  • ~13%: chip area OpenAI's models cut in the vector unit of its inference chip, shown in real silicon (OpenAI).
  • 57%: share of B2B buyer prompts to AI that are mid-funnel, against about 8% bottom-funnel (Kyle Poyar, Profound data).
  • 15M+: internal Deutsche Bank files in a leak published by about 50 news outlets (Inside Deutsche Bank consortium, FT).
  • 1,000+: engineers involved in fixing last week's Microsoft outage (Gergely Orosz, citing a Microsoft team lead).
  • 30+: partner plugins in ChatGPT for Sales at launch, including Salesforce, HubSpot and Clay (OpenAI).

What Happened

Models and compute

  • Google released Gemini 4 Neon, a fast model priced at $3 / $15 per million tokens. Google says it ranks third on the Artificial Analysis Intelligence Index (Google).
  • Microsoft AI released seven models at Ignite, including MAI Cyber 2 Flash, MAI Image 2.7 and MAI Realtime 1. Customers can now also tune a frontier MAI model and run it in their own Azure setup (Microsoft AI).
  • Google DeepMind opened Genie 4 to Google AI Ultra subscribers. Users can change a world mid-session by typing, and it now has sound (Google).
  • OpenAI says its models helped design chips. It cut area by about 13% and 10% in two parts of its own inference chip, and AMD has used some of the changes in a design sent for manufacturing (OpenAI).
  • Pika showed Enact, a research preview where you direct AI actors while the video plays. You can move actors, talk to them and steer the camera. It is not generally available (Pika).

Safety and policy

  • Two critical remote-code-execution bugs were found in NGINX's HTTP/2 code. Fixes are in 1.30.7 and 1.31.9; turning off HTTP/2 is the stopgap (Wiz).
  • Ruby fixed seven memory bugs in its regex engine and string case conversion, some shown to allow remote code execution. Upgrade to 4.0.10, 3.4.13 or 3.3.14 (Ruby).
  • Vals AI found frontier models usually helped pick targets in drone and satellite images, often including civilians. When a model refused, telling it the task was a game got around the refusal (Vals AI).
  • About 50 news outlets published the Inside Deutsche Bank leak, more than 15 million internal files on financial-crime checks and clients (Inside Deutsche Bank consortium, FT).
  • Democrats are probing the Justice Department's role in the OpenAI copyright fight, according to a letter that was still circulating (Axios).

Business moved

  • Palo Alto Networks agreed to buy Applied Compute for $4.5bn in stock. The startup, founded in 2025 by former OpenAI researchers, post-trains and serves open-weight models (Palo Alto Networks, WSJ).
  • KPMG UK ran more than 100 AI agents through a full trial audit. Its head of audit said they sometimes asked better questions than people. Partners still signed off, and a human had to step in once (Financial Times).
  • Isomorphic Labs raised a $5B Series C co-led by Greenoaks and Sequoia, with a16z and Alphabet (Isomorphic Labs).
  • Genesis AI raised a $500M Series A led by Khosla Ventures, Premji Invest and Eclipse, to deploy robots (TechCrunch).
  • GC AI passed $40M in annual recurring revenue, 36 months after it was founded, with more than 2,200 company customers (GC AI).

Platforms shifted

  • OpenAI launched ChatGPT for Sales, a sales plugin plus more than 30 partner plugins, for ChatGPT Business and Enterprise (OpenAI).
  • Last week's Microsoft outage came from a human-written bug about four months old, not AI-written code. That is according to a Microsoft team lead Gergely Orosz spoke to; the official postmortem is not out yet (Gergely Orosz).

Keep reading

The rest of this issue: Agents can now open your AWS account and bill your Claude plan. Google's new fast model costs $3 per million input tokens.

The tools, resources, reads and numbers are behind this point.

It costs an email address. The same address goes on Makersfuel, the MakersClaw briefing that runs Tuesday to Saturday, and every issue carries an unsubscribe link.

The address is used for the newsletter and nothing else. Privacy.

Get Makersfuel in your inbox

Makersfuel is the MakersClaw newsletter: a five-minute briefing for founders building with AI, with the tools, resources and reads worth saving, and what actually happened. Five mornings a week, Tuesday to Saturday.