Skip to content

Week 21 Recap: Cheaper Frontier Models, an Agent Marketplace, and 145 GitHub Stars

September 27, 2026

Opus 5.5 and GPT-6 Sol and Luna landed in the same 24 hours and both cut prices, Claude got a marketplace, open-weight models now run most tokens, and the week 39 board broke last week's tie with 917 upvotes. Week 21 on Nick Launches, plus what ten years of GitHub stars say about going solo.

Your weekly roundup of fresh launches shaping how we work and build.

Week 21 was a week where the ground moved under everyone building on AI, and the board noticed. On Tuesday, Anthropic shipped Claude Opus 5.5 at a lower price than Opus 5 and OpenAI answered the same day with GPT-6 Sol and Luna at half the price of the GPT-5.6 line. On Wednesday, Claude got a marketplace with more than 2,000 plugins and connectors on day one. By Friday all four new models, Grok 4.7 included, were sitting in the GitHub Copilot picker. If you build with an agent, the tools you were using on Monday cost less and can do more by Sunday. All of it is below, with the numbers that matter and none of the ones that do not.

On the board itself, the tie is broken. Last week cast exactly the same number of upvotes as the week before; this week the week 39 board cast 917, and the story is not the new paid launches. Three launches that went live on the last day of week 38 and finished it with two or three votes each spent this week climbing: OYYO AI went from 3 votes to 36 and took the board outright. Every landing page in this issue got a hand check before making the list.

Featured Launches 🚀

The Pro and Premium launches on this week's board, ordered as the week page has them, with each one's real board position, and every landing page checked by hand.

  1. OYYO AI logo OYYO AI: 36 upvotes and #1 on the board. Last issue it sat at #16 with 3 votes, having gone live on the last day of the week. Twelve times the votes in seven days. An AI workspace built in Dubai that puts local and cloud models, agents, tools, files, APIs, connectors and persistent memory in one environment, with a coordinator picking the right resource per task. Privacy, portability and controlled execution are architectural rather than settings. Featured through September 30.
  2. Spec0 logo Spec0: 25 upvotes, #2 on the board, and the top new launch of the week. AI coding assistants trust your API specs completely, even when they are out of date. Spec0 keeps every OpenAPI spec in one registry with owners, versions and the teams that depend on each API, flags breaking changes on every pull request, and serves Cursor and Claude Code just the operation they need over MCP, about 300 tokens instead of a whole spec pasted into the chat. Specs stay in your Git repo. Open-source CLI, free tier.
  3. Shotbase logo Shotbase: 23 upvotes, #3 on the board, up from 3 votes and #15 last issue. A screen capture workspace for macOS where the capture, the edit, the library and the share link are one workflow instead of four. Local AI does the organising and tagging, which is the part that decides whether you ever find the screenshot again. Still climbing, and featured through October 4.
  4. Riftrook logo Riftrook: 16 upvotes, #6 on the board. An AI game challenge platform: compete in AI-powered games, test models against each other, and see which one holds up when the rules are a game rather than a benchmark. The first of four new launches this week from Mark Zschiegner, who is in this newsletter for the third week running.
  5. Arena logo Arena: 13 upvotes, #8 on the board, up from 2 votes last issue. Products get matched head-to-head against others in their category and the community votes the winner through. Entering is free, winning a run makes you a Champion, and the social proof is a record of beating named competitors rather than a badge you bought. Featured through October 4.
  6. Build Mac App logo Build Mac App: 8 upvotes, #11 on the board, live since Friday. A template for selling paid macOS apps with the plumbing already done: license keys, notarized releases, auto-updates and payment processing, pre-configured and written so an AI agent can work inside it. You build the app, the template sells it. One-time purchase.
  7. IndieTools logo IndieTools: 7 upvotes, #12 on the board. A human-reviewed directory of independent SaaS products, apps, developer tools and the founders behind them. Every submission gets read before it goes up, which is rarer than it should be in this category. Freemium.
  8. Inkbound logo Inkbound: 6 upvotes, #13 on the board, live since Friday. A browser-based writing app for novelists: chapters, characters and world-building in one Codex, a style assistant that flags repetition and weak verbs as you type, entirely client-side, and a free-form timeline for chapters and date markers. Premium adds an AI that reads your manuscript and fills the Codex for you, plus corrected voice dictation. No AI writing, on purpose.
  9. Pavago logo Pavago: 6 upvotes, #14 on the board, in its second and final week. A searchable marketplace of pre-screened remote talent, built by recruiters for people who do not have time to recruit. Browsing 400k candidates is free, interviewing is $99 a month, and the success fee only lands after the contract is signed. Featured through September 29.
  10. Routebase logo Routebase: 6 upvotes, #15 on the board, second and final week. Your OpenAPI spec is the original and the docs portal, mock servers, contract tests and monitors are generated from it. Publish a change and Routebase compares your live API against the spec field by field. Worth reading next to Spec0 at #2: two launches, two weeks apart, both betting that the spec becomes the thing agents read first.
  11. Dialfyne logo Dialfyne: 6 upvotes, #16 on the board, second and final week. A sales platform that charges for usage instead of headcount. AI roleplay, a human-led power dialer and platform email all draw on one prepaid wallet, reps are free, and unused wallet money never expires.
  12. Timeburg logo Timeburg: 5 upvotes, #17 on the board, live since Wednesday. A free browser city builder on real OpenStreetMap maps. Pick any city on Earth, start from the year it was founded or from today, and watch plagues, wars and booms reshape it. Nothing to install, sign in with Google and build. Mark Zschiegner's second of the week.
  13. AISlopp logo AISlopp: 5 upvotes, #18 on the board, carried over from last week. One prompt turns into short AI video: characters, product clips and surreal verticals. Previews are free and deliberately glitched, clean exports are for subscribers. Featured through October 4.
  14. ShareMyGPU logo ShareMyGPU: 4 upvotes, #20 on the board, live since Friday. Rent GPUs by the hour from owners, or list your idle hardware and keep 75%. Owners set the hourly price, checkout runs through Stripe, and renters get secure access without a cloud account. Mark Zschiegner's third.
  15. FromTheClip logo FromTheClip: 3 upvotes, #25 on the board, carried over from last week. Upload a clip or a photo and it generates the scene before or after it, keeping the same face. First scene is free after email confirmation, packs start at $2.99. Featured through October 4.
  16. Free Fax Desk logo Free Fax Desk: 1 upvote, #26 on the board, live since Sunday. A fax comes in and it already knows who sent it and who should handle it; sending is drag, click, done, and the cover sheet writes itself. Built for the founder's own chiropractic office, open source, and free for yours. Works with Phone.com today. One day on the board so far, and the last-day lesson from last issue applies in full.
  17. JobsiteQuote logo JobsiteQuote: 1 upvote, #27 on the board, live since Friday. Contractors snap photos, pick prices, and send a good, better and best quote before they leave the jobsite. The customer signs from a link sent by text or email. Mark Zschiegner's fourth of the week, and the most un-AI product he has shipped, which is a compliment.

Launch Weekly Winners 🏆

The most upvoted free launches of the week. No paid placement here, just upvotes. Every one of these went live on Monday, and the top two out-voted thirteen of the seventeen featured launches above.

  1. PowerOutageHistory logo PowerOutageHistory: 15 upvotes, top of the free section. Twelve years of real county-level outage data, reliability grades, and generator recommendations. Look up your county or ZIP code and find out how often the lights actually go out before you buy the generator. A data product with no AI in the pitch, and the free winner of the week. Free.
  2. ScriptTimer logo ScriptTimer: 15 upvotes. Paste a script and get your exact speaking time at slow, normal or fast pace, no signup. Eight free tools in one: script timer, Shorts eligibility checker, podcast episode calculator, words-to-minutes, reading time, a WPM test and a presentation timer with slide calculator. No premium tier, no hidden features, and it says so.
  3. claim.town logo claim.town: 14 upvotes. Claim a block for $1 in a city of blocks, where each block is a real site and the closer to the centre you are, the higher you rank. It is the third "buy your rank" mechanic on this board in two weeks, after HighRank.lol last issue and outbid.to further down this week's free section, and the first one that looks like a place rather than a list.
  4. Cheapfax logo Cheapfax: 14 upvotes. Send a fax from your iPhone at 50 cents a page, no subscription, no minimum, and failed faxes refund themselves. Two fax products in one week, one paid and one free, in 2026. Somebody still needs to fax things, and both of these builders know exactly who.
  5. BuyMicroStartups logo BuyMicroStartups: 14 upvotes. A marketplace for buying and selling micro-startups with revenue verified through Stripe. No brokers, no commission, listings are free to browse and sellers talk to buyers directly. For everyone on this list who has a side project earning $200 a month they no longer want to maintain.
  6. Tokenry logo Tokenry: 14 upvotes. A no-code platform to create, airdrop and manage tokens across Solana and eight EVM networks. Your keys never leave your wallet, contracts are auto-verified, and mint, burn and ownership transfer sit in one dashboard after launch. Paid.
  7. ResuMakeAi logo ResuMakeAi: 13 upvotes. Matches your resume against live job listings, scores it, and shows which keywords are missing before you apply. Rewrites the resume and cover letter, checks them against real ATS rules, and preps you for the interview. Freemium.
  8. Human Benchmark logo Human Benchmark: 13 upvotes. Quick free tests for reaction time, aim, memory, focus and visual tracking. No signup, instant results, works on a phone. The kind of page people send each other at 11pm, which is a distribution strategy on its own.

Launch Builders to Watch 👀

Builders who made this week what it was. Worth a follow.

The week in models 🧠

Two frontier labs cut prices on the same Tuesday, and that is the headline for anyone paying per token.

Claude Opus 5.5 shipped on September 22 at $4 per million input tokens and $20 per million output, down from $5 and $25 for Opus 5, with cache reads at $0.20 instead of $0.50. Anthropic puts the real saving nearer 40% because it finishes tasks in fewer tokens and generates more than 30% faster. It performs at roughly the level of Fable 5.1 on most work and beats it on the agentic coding benchmarks, at less than half Fable's price. Claude Code made it the default Opus the same day. One detail from The New Stack worth knowing if you run agent chains: Opus 5.5 ships with the same safety classifiers as Fable, and a flagged cybersecurity or biology request is transparently rerouted to an older model. If a multi-step workflow starts behaving differently on one step, that is a place to look.

GPT-6 Sol and GPT-6 Luna landed from OpenAI the same day, positioned as faster, cheaper frontier intelligence at 50% lower API prices than the GPT-5.6 line, with improved prompt caching, a cache dashboard and prewarming. Sol is the capable one, Luna is the cheap one. On Vercel's gateway, Luna and Nano were already processing more than twice as many tokens as Astra and Sol for about a ninth of the spend, and that was before the price cut.

Grok 4.7 arrived on September 21 as a same-price upgrade built for multi-hour agent tasks. The New Stack's verdict was that it still fails most of them, which is at least honest reporting.

GitHub Copilot put all four in the picker by Friday: Opus 5.5 and Sol on Pro+ and up, Luna and Grok 4.7 from Pro. The Copilot app also got local sandboxing in public preview, which limits what an agent can touch on your disk, network and credentials. If you have ever watched an agent read a .env it had no business reading, that is the feature.

The bigger picture is in Vercel's September gateway index, which the Financial Times picked up this week. Open-weight models ran 56% of all tokens in August, up from 7% in December, and did it for 14% of the spend. Anthropic still took 64% of every dollar. The average price per token fell 23.2% in a month. AT&T told the FT that open models now handle 40% of its employees' AI requests, with coding costs down as much as 56% for a 2% dip in quality. For a solo builder the practical reading is simple: route the boring, high-volume work to something open or to Luna, keep Opus 5.5 or Astra for the step that has to be right, and re-run your evals this week because the price list you budgeted against in August is gone.

Claude Marketplace opened on September 23 with more than 2,000 connectors and plugins, plus a section for Claude-powered products from Cursor, Lovable, Snowflake and others, and a partner directory. Enterprise customers can pay for third-party products out of their committed Anthropic spend, and Anthropic takes no commission on those transactions. Two days later the plugin directory got a proper submission portal with auto-validation and review status. If you have built an MCP server or a skill for your own product, there is now a storefront for it with a procurement path that already exists inside the buyer's budget. That is the kind of distribution channel that rarely opens for small builders, and it opened this week.

GitHub signals 📡

What the trending list looked like on Friday, September 26, from the agents-radar daily snapshot. Star gains are for that single day.

  • paperclipai/paperclip, +2,109. "The open-source app everyone uses to manage agents at work": an org chart, a budget and a task board for a team of agents, MIT licensed, self-hosted, bring your own models. The biggest single-day gain on the list.
  • vectorize-io/hindsight, +1,653. Agent memory that learns across sessions. Together with mem0 at 66k stars and cognee at 31k, this is the loudest signal on the board: stateless chat is over, and memory is where the open-source effort is going.
  • google/ax, +1,379. Google's agentic orchestration runtime, in Go, with more than 1,500 stars on day one. Google, NVIDIA and Anthropic all shipped infrastructure this week rather than models, which tells you where they think the moat is.
  • dream-num/univer, +1,050. Spreadsheets, docs, slides and PDFs as one runtime for agents to operate. Agents that work inside an office suite rather than through an API.
  • mattpocock/skills +583, obra/superpowers +468, anthropics/skills +189. Skill files are still the fastest way to make an agent good at your workflow, and three of them are trending at once.
  • pbakaus/impeccable, +306. A design language for AI coding agents. This one is in my daily stack, more on that below.

On our own board, GitHub Highlights took 13 upvotes on a free listing for doing one new open-source repo a day, properly: what it does, why it showed up now, how it works, where it is weak, and every star count a dated snapshot rather than a live badge. If the list above is useful to you, that site is the daily version.

Worth your time 🔎

This week I went through every repository I have ever starred on GitHub, across two accounts and ten years, and wrote it up: 145 GitHub Stars, Two Accounts: From CTO to Solo Founder.

The short version. Seven years leading teams, nearly 21,000 contributions, one star. One year solo, 116 stars. I starred the least when I built the most, because once the stack is decided you stop shopping. And what I star changed completely: in 2016 it was a toast library and an image cache, things you import. In 2026 it is a headless browser my agents drive, a design language they read, and a folder of skills. Same problems, different reader.

The post has the 20 repos that actually made it into my daily stack, sorted into agents, browsers and scraping, design and UI, and SEO and growth, and which lists turned out to be bookmarks. Three of them, impeccable, Matt Pocock's skills and Anthropic's skills, are on the trending list above this week. Read it, and then tell me the oldest repo in your own stars. I am curious what it says about where you started.

Distribution notes 📣

Three things from the week that are about getting found rather than building.

Product Hunt's front page is editorial now. Every launch guide published this month says the same thing in different words: nothing appears on the homepage automatically, the editorial team hand-picks what gets Featured against four criteria (useful, novel, high craft, creative), comments and builder replies count for more than upvotes, and a non-Featured launch typically sees under a hundred visitors. If Product Hunt is on your calendar, plan it as a four-week project and bring your own audience, or treat it as one channel of many.

AI answers are citing manufactured pages. Trellner Research ran 380 software categories through Perplexity's models and found that 59.8% of 7,534 citations went to domains ranked worse than 100,000th by traffic, and that three sister sites had generated 215,128 templated "best software" pages between them. The report is from earlier this month and it hit the top of Hacker News. For you it means two things: your product page competing for an AI citation is competing against a template farm, and the fix is the boring one, a page that says plainly what the product is, who it is for and what it costs, in text a crawler can read. Site Audit checks the crawler half for free.

Your model choice is now a distribution decision. The Claude Marketplace above is a storefront. Copilot's model picker is a storefront. Vercel's gateway index is effectively a chart. Where your product sits relative to those, as a plugin, a connector, a skill, decides whether a buyer finds you inside the tool they already pay for. Spec0 at #2 on our board is a product built entirely on that bet.

What's happening on Launch 📊

This week

  • 📥 259 products submitted, up from 204
  • 🚀 101 went live in their launch window, 28 of them featured
  • ⬆️ 917 upvotes cast, up from 843 and the tie is broken
  • 👤 262 new builders joined, and the directory crossed 5,000 builders (5,002)
  • 📬 233 new subscribers on this newsletter, 3,408 on the list

The bigger picture

  • 🟢 1,987 live products on the directory, up from 1,870
  • 🔗 Domain rating held at 75
  • 🧰 26 site crawls in Site Audit this week, 3,685 pages read, 8,727 findings returned
  • 📣

The Monday effect 📅

Last issue had a section called the last-day problem: four featured launches went live on Sunday and finished the week on two or three votes each. This week is the other half of that story.

Three of those four are now at #1, #3 and #8 on the board. OYYO AI went from 3 votes to 36. Shotbase from 3 to 23. Arena from 2 to 13. Their featured windows carried them into a fresh week, they got the full seven days of people scrolling past them, and the votes followed. The fourth, AISlopp, went from 2 to 5, which is still more than double.

The free section says the same thing from the other side. Every one of the eight free winners this week went live on Monday. Not most of them. All eight.

So the rule from last week stands, with a footnote. If your date is open, take the Monday. If you launched on a Sunday, do not panic on Tuesday; the second week is where the featured window earns its money. And if you are launching on Monday, September 28, you have already done the hard part.

Quick ask 🙏

Twenty-eight featured launches, 917 upvotes, a top free launch at 15 and a top paid launch at 36. For once the top of the board is a paid launch, and it got there on votes, not placement: it started the week at #16 with three of them.

If something in this issue deserves eyes, open its page and upvote it. Ten seconds, real difference. Free Fax Desk and JobsiteQuote have one vote each and went live in the last two days; the Monday effect above tells you what happens next if a few of you click.

If you have a launch here, run a site audit. It is free, it takes under a minute, and this week's 26 crawls averaged 335 findings each. There will be something in yours.

Thanks for being part of Nick Launches. If you want the daily behind the scenes, including which launches climb and why, follow me on X: @nicklaunches (opens in a new tab).

Building something? Submit it. Listing is free and the backlink is permanent.

21 weeks down. Thanks for reading. 🙏

Category: Product Updates
Nick LaunchesChangelogRecapProduct UpdatesLaunches
Share

The weekly launch recap, every Sunday.

What launched, what won, and what the top of the board did differently.

By submitting your email you agree to our Privacy Policy