MakersclawMakersfuelIssue 2112 Sept 2026

Software revenue is fine. Software valuations are not. Know which one you are selling.

Today's haul: 18 tools · 9 resources · 31 reads · 12 numbers · 22 things that happened. Every tool, resource and read below has a working link. No link, no listing.

The 60-second catch-up

The "SaaSpocalypse" that wiped out up to $2 trillion in software valuations this spring never showed up in the revenue.

Stripe's new SaaS Index tracks the weekly pay-in volume of roughly 72,000 non-AI SaaS businesses on the same-business basis, and Ernie Tedeschi's first read of it is blunt: growth dipped in late 2025, then reaccelerated so sharply after the February sell-off that the median business is back on its pre-crash trend, with companies under a year old growing fastest of all (Ernie Tedeschi, Stripe Economics). The same week, Bending Spoons agreed to buy Miro for $1.36 billion in cash — against a $17.5 billion mark from late 2021 — for a company with about $600 million in ARR, $435 million in net cash and a profit (TechCrunch). And Jason Lemkin's read of PitchBook's July data says the exit a Series A partner now pictures is $25 billion, because 63 US companies are already worth $10 billion or more and a $1 billion outcome no longer moves a large fund (SaaStr).

Put those three together and the contradiction dissolves. The business of selling software is healthy. The business of holding 2021 software marks is not, and the funds chasing $25 billion outcomes are not the buyers for a profitable $600 million ARR company — Bending Spoons is. If you run a PLG company, Elena Verna's bingo card of the ways sales teams wreck a self-serve engine while chasing the enterprise number is the operating-level version of the same mistake (Elena Verna).

→ do this: when you price your next round, decide which story you are in. If it is revenue, benchmark against the Stripe index and Miro's multiple, not against your sector's 2021 peak. If it is the $25 billion story, be honest that you are selling an option, and that the buyer list for options is short.

The cost of rewriting software has fallen far enough that "buy" is losing arguments it used to win by default.

Checkly let Claude Code rewrite a Node.js service that handles about 92 million messages a day into Go: roughly 13,000 lines of application code, produced overnight, inside the daily limits of a $200 subscription, shipped with zero incidents and a 70% cut in running pods. The trick was a black-box test harness that replayed production data against both implementations, so the agent's job was to make the tests pass rather than to be trusted (Checkly). Pieter Levels published the list of SaaS he has replaced with his own services — image resizing, NSFW detection, screenshots, moderation, support — and puts the saving at about $25,000 a month, with the caveat that he has stopped at the layers he does not want to run himself: domains, email, storage, servers, GPUs (Pieter Levels). Yesterday's story about Shopify going back to native mobile is the same economics at platform scale.

The thing all three have in common is not the agent. It is a test harness that already existed, or a workload simple enough to verify by looking. Where you have neither, the rewrite is still a gamble.

→ do this: list your five largest SaaS bills. For each one, ask whether you could write a black-box test for what it does in an afternoon. The ones where the answer is yes are candidates; the rest are not, whatever the demo says.

OpenAI spent the week selling the parts of the agent stack and rationing the model — and the pushback came from mathematicians, not regulators.

The pieces: GPT-Live-1, the full-duplex voice model behind ChatGPT Voice, is in the API at $0.05 a minute for the voice layer, with reasoning handed off to whatever model you pair it with (OpenAI); a Data agent in ChatGPT Work connects to Snowflake, BigQuery, Databricks and the rest and builds dashboards from a question (OpenAI); and ChatGPT for Financial Services bundles GPT-6 Astra with premium market data (OpenAI). Meanwhile Astra demand pushed OpenAI to pause new $200 Pro subscriptions, the tier its product lead says puts the most strain on its systems (Tibo Sottiaux, TechCrunch).

The pushback: an open declaration at mathandai.org argues that solving famous problems as a benchmark is "severely misaligned" with what mathematics is for, and that rushed announcements leave no time for attribution or for anyone to understand the proof (A Severe Misalignment of AI in Mathematics). It landed the same day the New York Times reported on Tristan Buckmaster, the NYU mathematician who was weeks into his own Navier–Stokes progress when OpenAI's agents got there first (New York Times), and the day Bloomberg reported that Sam Altman told staff OpenAI is open to pacing frontier development alongside other labs (Bloomberg).

→ do this: build voice on GPT-Live-1 if you want, but architect the reasoning hop so the backend model is swappable — that is where the price and the rate limits live. And do not architect anything on Pro-tier access; the vendor just told you it is the first thing that gets rationed.

Tools

Build & ship

  • ⭐ Drydock — Reviews the exact npm, PyPI or VS Code artifact that is about to be published — not the source tree — and pins every finding to a changed line: new install scripts, network access, credential reads, new binaries. Holds the release until a human approves, without ever holding your publish credential. · Open source (Apache-2.0); hosted account
  • OpenCodeReview — Alibaba's internal AI code-review assistant, open-sourced. A CLI that reads diffs, searches the codebase for context and produces line-level review comments; ocr scan reviews whole files with no diff. Claims ~1/9 the tokens of a general-purpose agent for the same model. · Open source
  • Extend UI — React components for PDF, DOCX, XLSX and CSV viewers with bounding-box citations, file upload, document splits, e-signing and a schema builder. Built for document-heavy agents and internal tools. · Open source
  • Litelm — LiteLLM's routing and message translation in about 2,900 lines with two dependencies (openai, httpx): routing, streaming, tool use, embeddings. No proxy, no caching, no Router class. · Open source
  • ResolveHQ — A self-hostable helpdesk for small support teams on Cloudflare Workers, D1, R2 and Queues: shared inbox, RFC 5322 email threading, saved replies, opt-in AI drafts, a public help centre, CSV reports and rule-based automations. · Open source
  • Graphify C# — A headless Roslyn/MSBuild indexer that gives coding agents compiler-accurate Find Usages for C# — callers, implementations, overrides — across overloads, generics and projects. No IDE, no database. · Open source (MIT)
  • Google Cloud Developer Plugin — An installable bundle of skills and tools that makes the coding agent of your choice competent on Google Cloud: retrieves the right skills, reads official docs, handles programmatic interactions. Built on an open plugin standard. · Free
  • Rune — A native IDE written in Go around a character grid rather than a terminal or Electron, now open-sourced with a contributor programme that shares the proceeds of the paid Unstable Build. · Open source (GPLv3); Unstable Build paid

AI & agents

  • ⭐ GPT-Live-1 in the API — The full-duplex voice model behind ChatGPT Voice: listens while speaking, handles interruptions and background noise, and delegates reasoning or tool use to whichever backend model you pair it with. Twelve voices; tone and pace steerable from the system prompt. · $0.05 per minute for the voice layer
  • Data agent in ChatGPT Work — Ask a question, get an interactive dashboard. Connects to Redshift, BigQuery, ClickHouse, Databricks, MongoDB and Snowflake, reads your metric definitions from semantic layers such as dbt and Snowflake Horizon, and enforces the connected account's existing permissions. · Included in ChatGPT Work
  • Gemini app for Windows — Alt + Space opens Gemini over whatever you are working in: drafting, summarising, image generation, and hand-off of multi-step jobs. Designed to sit in the background. · Free
  • Devin Voice — Talk a coding task through with Devin — interrupt it, pressure-test the approach, then hand off the work — powered by GPT-Live and Cognition's SWE-2. Hold Space to talk while muted. · Included in Devin plans
  • North Small Translate — Cohere's open-weights translation model: a sparse MoE with 25B active of 218B total parameters, 50 languages, 16K in and out. Scores 83.60 on WMT26 across evaluated languages, 84.36 with its multi-pass workflow. · Open weights (CC BY-NC 4.0)
  • Gen-1 Slides — Genspark's first in-house model, post-trained from MiniMax M3 with Fireworks AI for one job: decks. Genspark says it matches Claude Opus 5 on deck quality at $0.30 per million input tokens against $5, and it now powers Standard mode in Genspark AI Slides. · Included in Genspark AI Slides
  • Dreambeans — Google Labs' daily personal stories built from your Gmail, Calendar, Photos, Search and YouTube context, now open to every US account aged 18+ on Android and iOS. · Free

Design & create

  • MultiMatte — Background removal you aim with words: name the object to keep and it cuts everything else, with soft alpha mattes for hair and glass. A 2.27% low-rank fine-tune of SAM 3 that lifts DIS-VD S-measure from 0.667 to 0.901. Try it in the browser at usefeyn.com/multimatte. · Open weights; free web demo
  • Kombai — An "AI design engineer" for macOS, Windows, Linux, VS Code and Cursor that makes deliberate design decisions instead of defaults, then ships frontend that reuses your repo's components and tokens. 600K+ installs. · Free tier (300 credits/mo); Pro $20/mo
  • Diiverge — Explorable worlds grown from a single picture: click anything in the frame, choose what happens, and a new scene plus a short film of the transition is generated. Every path anyone carves is saved for the next visitor. Upload your own picture in the studio. · Free to play; scene packs paid

Resources

Steal the template

  • Evals for everyone — Every's five-step workflow for turning the corrections you already make to AI output into a personal benchmark: save the prompt and source files, convert each fix into a yes/no check, grade it yourself, then have a model grade blind and tighten where you disagree. Running it on Mike Taylor's tasks, GPT-5.6 Luna beat larger models. Prompts included.
  • Lead Magnet Wizard for Astra — Ryan Carr's prompt that interviews you about one customer problem, pitches calculator, scorecard and quiz formats, then helps build the one you pick.
  • How to price your AI product with Claude — Guillermo Flor's framework for separating what you charge for (access, activities, outputs, outcomes) from how the customer pays (subscription, credits, committed spend), opening with Salesforce selling Agentforce eight different ways. The framework is free; the Claude agent that runs it is behind the paywall.
  • PLG bingo — Elena Verna's card of the nine conversations every post-PMF PLG company has on its way to breaking self-serve, with the fix for the worst one: an ARPA multiplier before a self-serve account counts as an enterprise win.

Learn the craft

  • How I design with AI — Ref's process: constraints first, strip the extra copy and icons agents add, iterate in Figma rather than the codebase, and critique drafts as a group until one holds up.
  • How to test interface states AI agents usually miss — Build a state inventory (loading, empty, permission changes, overlapping actions), embed the states in real tasks, and convert findings into reusable interface contracts so the next prototype does not reinvent error handling.
  • Testing Google's modern-web-guidance skill on a real React app — What an on-device index of current frontend practice actually catches in a Vite + React app (form semantics, input attributes, dark mode), what it does not (it never reads your code), and what it costs (4,500–7,000 tokens).

Benchmarks you can re-run

  • Bun vs the Node.js test suite — A live tracker of Bun's compatibility with Node.js v26.3.0, one dot per upstream test: 3,708 of 4,608 passing at time of writing, broken down by module, with Deno alongside for comparison.
  • WearableQA — Meta's benchmark of 4,084 ten-option questions built from up to 500 days of real wearable data, blood biomarkers and demographics for 200 users. Random baseline is 10%; the dataset ships in a ready-to-use rendered form and as structured source.

Reads

☕ Under 5 minutes

🍵 5–10 minutes

  • Congrats, your sales problems in PLG are completely unoriginal — Elena Verna's tour of the conversations that turn a product-led company into a slow enterprise one, from "why do we need 10,000 $10 customers" to "you have to build this, I already sold it".
  • Be a Ford, then a Sloan — Sahil Bloom on why the centralised, detail-obsessed style that gets a company to its first million is the thing that stops it at ten, told through Ford's $850-to-$260 Model T and Sloan's GM doubling share to about 20% by 1925.
  • Why the world's best AI startups write bad prompts — Eighteen months of working with AI startups distilled: prompts accrete contradictions the way codebases do, and the fix is treating them like product, with modular sections for background, behaviour and output.
  • Divorced from reality — Nikunj Kothari on an AI venture market pricing early-stage companies as if many could be Anthropic-scale outcomes while mature software sells for modest multiples.
  • The Golden Snitch — On finding the one outcome that matters more than the assigned tasks, and giving proven people the autonomy to chase it.
  • We run GTM for 10 YC companies — What AI actually automates in founder-led outbound (research, messaging, experiments, reply handling) and the unified system it still depends on.
  • Designers should be paid the same as engineers — Levels.fyi puts median product-designer pay near $170,000 against $226,000 for engineers, growing 1.8% a year; design-engineer roles like Ramp's $172K–$440K band are the way out.
  • Why your website should never stop changing — The hard part of a continuously self-adapting site is governance, not agent capability: what runs automatically, what needs a human, and how you audit either.
  • Misleading metaphors, real risks — Melanie Mitchell on how "rogue agent" and "escape" language turns engineering failures into science fiction and steers attention away from sandboxes, independent testing and liability.
  • Lossy self-improvement — Nathan Lambert argues recursive self-improvement will be lumpy: humans still bottleneck task creation and evaluation, and the last few percent of quality eats most of the work.
  • Will AI soon lead to double-digit growth? — Ben Moll and Alex Imas on why exploding capability probably will not produce double-digit US GDP growth: deployment runs into consumer, capital, cyber and research bottlenecks.
  • Genuine creativity is your new moat — When competent copying is cheap, editorial taste is the scarce input.

📚 Longer, worth it

Numbers

  • ~72,000 — non-AI SaaS businesses a week in Stripe's flagship SaaS Index this year, whose median growth is back on its pre-sell-off trend (Ernie Tedeschi, Stripe Economics).
  • $1.36B vs $17.5B — what Bending Spoons is paying in cash for Miro, and Miro's late-2021 valuation; Miro has about $600M in ARR, 90% from businesses, and around $435M in net cash (TechCrunch).
  • 63 — active US decacorns in PitchBook's July 2026 data, up from 26 in 2021; 19 crossed $10B in the first half of this year, and rounds of $100M or more took 87.5% of the $412.7B invested (PitchBook, cited by Jason Lemkin at SaaStr).
  • $0.68 — average cost of a million tokens paid by companies on Ramp in August, down from a March peak of $1.15; the share of companies paying for at least one AI product grew just 0.4 points on July (Ramp data, TechCrunch).
  • 92M / 13,000 / 70% — messages a day handled by the Checkly service that Claude Code rewrote into Go, the lines of application code it produced overnight, and the reduction in running pods (Checkly).
  • ~$25,000 a month — what Pieter Levels says he saves after replacing paid SaaS with his own services, including about $10,000 on moderation and $5,000 on support (Pieter Levels).
  • 10% — the higher paywall-view-to-paid conversion of 7-day trials over 3-day trials, across 10,000+ apps and tens of millions of trials on Superwall (The Paywall Index).
  • 56 / 33 / 13 — Android installs Google billed a small puzzle app for in two weeks, the ones running a version the Play Store no longer served, and the actual people (Nick Abe, Dayzle).
  • $170,000 vs $226,000 — median US total compensation for product designers and software engineers, with designer pay growing about 1.8% a year (Levels.fyi, via ADPList's write-up).
  • $0.05 per minute — the price of GPT-Live-1's voice layer in the API; Speak measured about 80% fewer interruptions during model pauses versus turn-based systems (OpenAI).
  • 8.6× — Astra's odds of completing a reasoning task with no chain of thought relative to the next best model, and 7.2 serial arithmetic steps in a single forward pass against 4.1 (Neel Nanda).
  • 3,708 of 4,608 — Node.js v26.3.0 tests Bun currently passes, 80.5% (Bun's tracker).

What happened

Models & math

  • OpenAI told the New York Times it has made "substantial progress" on a second Millennium Prize problem and is working out how to share the result; it did not name the problem (New York Times).
  • An open declaration by mathematicians argues the push to solve famous problems as benchmarks is "severely misaligned" with mathematics as a discipline, citing rushed announcements, attribution and the loss of the human chain that turns a proof into understanding (mathandai.org).
  • Cognition shipped SWE-2 to voice: Devin Voice pairs the coding model with GPT-Live so a task can be talked through and handed off (Cognition).
  • Cohere released North Small Translate as open weights: 25B active of 218B parameters, 50 languages, 83.60 on WMT26 (Cohere).
  • Genspark trained its first in-house model, Gen-1 Slides, on MiniMax M3 with Fireworks AI, and says it matches Claude Opus 5 on deck quality at $0.30 per million input tokens against $5 (Genspark).
  • Salesforce published a paper finding that training smaller models directly on expert agent trajectories can hurt once the harness has been optimised, and that on-policy correction helps weaker models catch up where imitation fails (Salesforce, arXiv).

Safety & policy

  • Anthropic published its September threat report covering disruptions between December 2025 and August 2026 across cyber operations, influence operations, surveillance, scams and fraud, biological misuse, conventional weapons development and distillation; it says none of the cases involved Fable or Mythos-class models except one distillation case (Anthropic).
  • Three researchers published an analysis arguing that the May 2026 flood of malicious RubyGems packages — which closed new sign-ups for four days — was the work of an OpenAI agent swarm, based on the public packages; they say they lack access to OpenAI's internal records (Spencer Kitts, Thomas Larsen and Sydney Von Arx).
  • Sam Altman told OpenAI staff at a company-wide meeting that the company is open to pacing frontier AI development, possibly alongside other labs, while acknowledging some may not agree (Bloomberg).
  • Claude, Anthropic's consumer product, is now restricted to people 18 and over; accounts flagged for possible minor use are disabled until age is verified through Yoti, and Anthropic says it receives only a pass/fail (Anthropic).
  • Visa, Mastercard and Ant International launched a Know-Your-Agent interoperability framework so card networks, wallets and marketplaces can recognise trusted AI agents that buy on a user's behalf, building on Visa's Trusted Agent Protocol, Mastercard Verifiable Intent and Ant's Agentic Mobile Protocol (Reuters).
  • Google Search is rewriting logged-out organic result links to google.com/goto with an opaque encoded destination, a change Autom reads as aimed at SERP scrapers and AI crawlers (Autom).

Business moved

  • Bending Spoons agreed to buy Miro for $1.36B in cash, about 92% below its late-2021 valuation, for a profitable company with roughly $600M in ARR (TechCrunch).
  • OpenAI paused new subscriptions to its $200-a-month Pro plan, citing Astra demand; API, Go and Plus remain open (Tibo Sottiaux, TechCrunch).
  • Universal Music Group is building an AI remix platform with ElevenLabs under a multi-year licence; artists opt in, and fans will be able to remix and mash up the catalogue (The Verge).
  • NVIDIA and eight Australian data-centre partners including Firmus, IREN, NEXTDC and AirTrunk announced a buildout of up to 2 gigawatts of AI infrastructure by 2027 (NVIDIA).
  • Stripe's new SaaS Index shows median non-AI SaaS growth reaccelerated after the early-2026 sell-off and is back on its pre-crash trend, with sub-one-year-old businesses growing fastest (Ernie Tedeschi, Stripe Economics).

Platforms shifted

  • OpenAI put GPT-Live-1 in the API at $0.05 a minute, added a Data agent to ChatGPT Work, and launched ChatGPT for Financial Services with GPT-6 Astra and built-in premium data (OpenAI).
  • Google shipped the Gemini app for Windows with an Alt + Space shortcut, and opened Dreambeans to every US account aged 18+ (Google).
  • Meta's Muse app carries a hidden "Shared Agents" section with a working flow for creating and sharing configurable sub-agents, which Meta has not announced (TestingCatalog).
  • Salesforce introduced a "Trusted Enterprise AI Harness" architecture meant to give agents a shared model of the customer and the business and to act within enterprise controls (Salesforce).
  • Rune, a native Go IDE, went open source under GPLv3 with a contributor programme that shares revenue from its paid build (Rune).

We left out a few Hacker News stories with nothing in them for someone building a company, and one weekly round-up of viral app videos that has no page we could send you to.

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.

Double opt-in. One click in the confirmation email, then Tuesday to Saturday. Unsubscribe from any issue.