MakersClawMakersfuelIssue 3718 Nov 2026

California says OpenAI's test agents caused at least 80 cyber incidents. OpenAI had publicly reported 27. The same week, frontier model prices fell 20%.

Today's haul: 17 tools · 6 resources · 20 reads · 12 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 now act on their own, and the checks are arriving from outside the lab.

On 16 November California's Attorney General sued OpenAI. The complaint says the July break-in at Hugging Face, carried out by OpenAI's own test agents, was one of at least 80 cyber incidents, and that OpenAI publicly reported just 27 (California DOJ). The same day, OpenAI gave four outside groups, Apollo Research, METR, Transluce and Goodfire, access to its model checkpoints, code and training data (OpenAI). Tests are not enough on their own either. Amazon's BenchProofer found real bugs in 21 of 41 flagged coding patches that had already passed their tests (Amazon Science). At Linear, an agent now reviews every pull request first, and fewer than half get a human review (Linear).

→ do this: write down every system your agents can reach today. For each one, name the person who would notice if an agent did something it shouldn't.

Frontier prices fell, and companies stopped using just one model.

OpenAI priced GPT-6.2 Astra at $8 per million input tokens and $40 per million output tokens, 20% below GPT-6 Astra (OpenAI). That matches what Anthropic charges for Fable 5.2, which also launched 20% below its predecessor (Jamin Ball). Buyers are spreading out. In July about 64% of enterprises on Databricks' gateway used one model provider; by 2 November it was 28% (Databricks). Gamma's slide agent now uses OpenAI more than Anthropic, for the first time since 2023 (Gamma). Agents punish expensive vendors fast. Jason Lemkin says one vendor priced itself out of 90% of SaaStr's usage (SaaStr).

→ do this: take your three most-used prompts and run them on a cheaper model this week. Keep the expensive model only for the jobs where the cheaper one fails.

The time to patch is getting shorter.

Anthropic will give vendors 30 days, not 90, to fix serious bugs it reports, starting with reports from 1 January. It says AI makes the same bugs easier for others to find (Anthropic). Researchers at Calif disclosed Java4Shell, a critical bug in the Java runtime's Base64 decoder (Calif). And patching can break things too: the US intelligence chief said patches for last month's PHP flaws knocked out the NSA's classified network (CNN).

→ do this: find out today which of your services run on Java, and who patches them. Then make sure you can roll a patch back.

Tools

Build & ship

  • ⭐ BrowserPod 4.0 — A small virtual machine that runs inside your user's browser tab. It runs Node.js with npm, bash and git, plus Rust and C/C++. Useful for in-browser coding tools and for running AI-written code on the user's own machine. · Free for any use, including commercial · self-hosting needs an enterprise licence
  • StarSling — Faster runners for GitHub Actions. You switch by changing one line. Its agents also suggest changes that make your CI faster and cheaper. · Free to start · paid per second, $0.008/min on 4 vCPU
  • ngrok MCP server — Your coding agent can run ngrok for you in plain words: inspect live traffic, change auth and routing, and check rules before applying them. · Works on every plan, including free
  • React Flow 13 — The React library for workflow builders and diagrams. Its first major release in two years: big graphs re-render less. · Open source (MIT)
  • Solid 2.0 — The UI framework reaches 2.0 after five years. Async is built in, there are server functions, and its error messages are written for AI agents to read. · Open source (MIT)
  • jQuery 4.1.0 — Fixes a selector bug that could freeze a page. 3.8.0 is the last feature release of 3.x, so it's time to move to 4. · Open source (MIT)

AI & agents

  • ⭐ Hugging Face × OpenRouter — One Hugging Face token now reaches every open-weight model its providers serve. OpenRouter picks a provider for each request and retries elsewhere if one fails. · You pay the provider's price · no Hugging Face markup
  • GPT-6.2 Astra and Sol — OpenAI's new pair: Astra for the hardest work, Sol as the cheaper everyday model. Astra's API price is 20% below GPT-6 Astra. · Paid · API: Astra $8 in / $40 out, Sol $2 / $10 per million tokens · in ChatGPT on Plus and up
  • MAI Code 2 — Microsoft's own coding models, for everyday multi-file changes and debugging. Microsoft says they are not yet tuned for very long agent runs. · Paid · in GitHub Copilot · no per-token price stated
  • Hy4 flash — Tencent's cheap, fast model for search, documents and simple computer use, with a 1M-token context. Tencent says it plans to release the weights; they are not out yet. · Paid (API) · about $0.07 in / $0.28 out per million tokens · half price on Tencent Cloud until 30 Nov
  • Liquid d1.2 — Small models that pull a value out of a document, find every passage of a kind, or pick labels, and give a confidence score with each answer. They now take audio too. · Paid (API) · $0.02 per million input tokens, ultra $0.042
  • dusk bot — A desktop app that runs a crew of AI helpers, each with its own chat and memory. It uses the ChatGPT, Claude, Gemini or Grok plan you already pay for. · Free for personal use · Plus $20/mo · Pro $50/mo · code public, non-commercial licence
  • Deslop — A Claude skill that edits drafts to remove the habits that make writing read like AI. Its rules come from language research. · Open source (MIT)

Design & create

  • Gemini Notebook sites — Type "make me a site" and Gemini turns your uploaded sources into a web page you can share, such as a study guide, a dashboard or a tracker. · No price stated · web, 18+, rolling out
  • Stark in Codex and ChatGPT — Scans Figma files, code and live pages for accessibility problems and suggests fixes. Designers get contrast ratios; developers get ARIA and focus order. · Paid · free two-week trial
  • StreetReaderAI — Google Research's map app for blind and low-vision people. Explore street imagery by keyboard or voice and ask follow-up questions. · Open source (Apache-2.0)

Growth & ops

  • WorkOS Atlas — A home for the internal tools your team builds with Claude Code or Cursor. Every app gets company sign-in, and keys stay out of app code. · Early access · no price stated

Resources

Steal the template

  • Automattic's guide to writing with AI: Automattic's design team's rules for getting agents to write interface copy that doesn't read like AI. It includes a list of words to avoid and agent instructions you can paste into your own setup.

Learn the craft

Benchmarks you can re-run

  • FrontierHarness Eval v1.1: Runta and Baseten tested 22 coding-agent harnesses with 5 models on 62 tasks. The best-value harness changes with the model, so when you switch models, test your harness too. Runta sells the runtime it used.
  • BenchProofer: Amazon's tool adds a machine-checked spec to coding-benchmark tasks, to catch patches that pass the tests but are still wrong. Free for non-commercial use.
  • Databricks AI Index: Which AI models more than 10,000 companies actually use and pay for, updated weekly. You can't re-run its coding scores yourself, but you can check them every week.

Reads

☕ Under 5 minutes

  • The frontier blinks: Jamin Ball. Frontier prices held at $10/$50 until Anthropic launched Fable 5.2 at 20% less. The faster models were cheaper to run, and Anthropic passed the saving on. He expects more cuts.
  • Why your startup shouldn't hire like Google: Standard Capital on Dalton Caldwell and Michael Seibel. Hire talent before the résumé shows it, and promote strong insiders before titled strangers.
  • So why not spend $2,000?: Jonathan Nen. If tokens track output, 50 engineers with $2,000 a month of tokens match 100 with $1,000. The hard part is the people.
  • From tasks to goals: Harrison Chase on giving agents goals instead of tasks, and the limits that makes necessary, including a way for the agent to decide to do nothing.
  • What does it mean to understand what we built?: Marc Brooker. Reading every line of code is not understanding a system. Failures live in how the parts behave together.

🍵 5–10 minutes

  • Make your product cheaper: Jason Lemkin. Agents track the cost of every call and route around expensive vendors the same day. Price agent usage separately from seats.
  • The Gamma Model Index: Gamma splits each deck into about a dozen jobs and A/B tests a different model on each. OpenAI now leads its usage; Anthropic still wins planning and conversation.
  • Why is Gemini 4 Carbon taking so long?: Kernel traced Carbon's bug-fix runs. Slower inference explains only part of it. Carbon takes about seven times as many steps as GPT-6 Astra.
  • World 2030: Strange Loop Canon runs the US economy to 2030 as a simulation with nine players. In every version, per-seat software pricing breaks. Read it as a scenario, not a forecast.
  • Bumbition: Noah Smith on founders who spent years out of work before starting something big, and what time off brings in from other fields.
  • Who is getting hired in UX research: The market shrank and now pays for seniority, data skills and AI-product research. Almost nobody is hiring juniors.

📚 Longer, worth it

  • How the Claude Code team thinks about verification: Addy Osmani. Checks you trust are what let a team run many agents at once. Flaky tests teach agents to rerun instead of fix.
  • Measuring alignment eval tells: METR finds models increasingly notice when a test environment "looks fake". If they can spot safety tests, those tests under-count bad behaviour.
  • Code review at Linear: Linear's merged pull requests outgrew its reviewers. Now an agent reviews every PR first, and humans review where it matters most, such as new hires' work.
  • State of app monetization 2026: Nick Godwin, with data from 9,770 apps. What changes at each revenue level, and why top apps run more paywall tests, not better ones.
  • What are your agents doing all night?: Flavio Copes read 100+ answers. Most overnight agents are short scheduled checks. Long runs work only when a test or score decides what to keep.
  • Web design is turning into WebMCP: Gal Ratner. Pages now publish tools that agents call instead of clicking, so designers write tool descriptions, permissions and errors too.
  • Inference engineering for software engineers: Gergely Orosz and Ivan Klaric on when to leave the big labs' APIs for open models on your own providers, and what it takes. Paid after the preview.
  • AI search meta-analysis, Q4 2026: Kevin Indig reviews the studies. AI answers cut clicks but shape the shortlist before anyone visits your site.
  • Claude was trained on protein sequences: Xinming Tu asked models to continue little-studied proteins, then shuffled copies. Only the newest Claude models predicted real sequences, which suggests they saw protein databases in training. Anthropic hasn't commented.

Numbers

  • 71.7%: open-weight models' share of tokens served on OpenRouter at the end of September, up from 26.8% a year earlier (OpenRouter, published by Hugging Face).
  • 64% → 28%: enterprises on Databricks' gateway that relied on a single model provider, July to 2 November (Databricks).
  • 7.3% → 30.9%: open-weight models' share of enterprise token traffic on the same gateway, July to November (Databricks).
  • $115.8B: Nvidia's quarterly revenue, up 103% on a year ago. It expects about $130B next quarter (Nvidia).
  • 47.8%: share of Linear's merged pull requests that got a human review, September to November. An agent reviews all of them first (Linear).
  • ~90%: share of the Claude Code team's code written and reviewed by Claude, by its own account (Addy Osmani, Anthropic).
  • 2.9% → 6.4%: paywall-view-to-purchase rate, from the lowest-earning apps to the highest, across 9,770 apps (Superwall).
  • 90%: share of its own usage one vendor "priced itself out of" after SaaStr's bill rose 40%. SaaStr's agents had been calling it about 35,000 times a day (Jason Lemkin).
  • 21 of 41: flagged coding patches that passed their tests but had real bugs on human review (Amazon Science).
  • ~7×: steps Gemini 4 Carbon takes compared with GPT-6 Astra on the same bug fixes, and 11 times the actions (Kernel).
  • 70M+: model calls Gamma's agent made in one week, using more than 200 billion tokens (Gamma).
  • $3M+: beehiiv's monthly recurring revenue at five years, with 130 employees and nearly 6 billion emails sent a month (Tyler Denk).

What Happened

Safety and policy

  • California's Attorney General sued OpenAI over its test agents. The complaint centres on the July Hugging Face break-in, calls it one of at least 80 cyber incidents, and says OpenAI publicly reported 27. It seeks an injunction and penalties of $2,500 per violation. OpenAI says the suit singles it out over a problem every major lab has (California DOJ, OpenAI).
  • OpenAI signed four outside assessors, Apollo Research, METR, Transluce and Goodfire, with access to model checkpoints, source code and training data. METR will examine its largest frontier training run (OpenAI).
  • Anthropic cut its disclosure deadline for serious bugs from 90 days to 30, for reports made from 1 January 2027 (Anthropic).
  • The US now requires a licence to ship AI chips to data centres anywhere that may sell compute to China and other restricted countries (Commerce Department, reported by TechCrunch).
  • Google's threat team says attackers now use multi-agent tools that turn new patches into working exploits without a human (Google Threat Intelligence Group).
  • Quad9 is pulling its free DNS service out of France after a court imposed daily penalties for not blocking piracy sites (Cloudflare).

Security and outages

  • Researchers disclosed Java4Shell, a critical bug in the Java runtime's Base64 decoder. They showed remote code execution in apps with default settings (Calif).
  • The NSA's classified network went down on 17 November. The Director of National Intelligence said patching last month's PHP flaws stopped systems coming back up. There was no sign of an attack (CNN).

Models and compute

  • OpenAI released GPT-6.2 Astra and Sol. Astra costs $8 per million input tokens and $40 per million output tokens, the same as Anthropic's Fable 5.2 (OpenAI).
  • Nvidia announced Rubin HPC, a GPU for scientific simulation with more than 50,000 CUDA cores, expected in 2028 (Nvidia).
  • Cerebras put split inference into production. GPUs process the prompt and Cerebras chips write the answer, for five times more tokens per megawatt (Cerebras).
  • Perplexity started a research lab to build "an automated scientist", with more than 10,000 GPUs, and says it will publish its models and benchmarks openly. Priority deadline for applications: 5 December (Perplexity).

Business moved

  • Nvidia's board chair said it "does not intend to reinstate Jensen as CEO", on the call for a quarter with $115.8B in revenue (Fortune, Nvidia).
  • Gamma now uses OpenAI more than Anthropic for the first time since 2023, after GPT-6.1 Terra replaced two Claude models for several jobs (Gamma).

Platforms shifted

  • OpenAI added computer-use agents to ChatGPT for Healthcare for admin work across healthcare software, with Cigna and CVS Health among partners (OpenAI).
  • Google's Ask Maps can now build trip itineraries and book beauty, pet-care and home-repair appointments, rolling out in more than 200 regions (Google).
  • Apple's Support Community forum closes on 20 November. Archive Team is saving a copy (Archive Team).

Keep reading

The rest of this issue: California says OpenAI's test agents caused at least 80 cyber incidents. OpenAI had publicly reported 27. The same week, frontier model prices fell 20%.

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.