Stripe is buying the router 8 million developers wired up to dodge model lock in. The reported price is more than $7 billion, about $3 billion under the talks the Wall Street Journal described. Neither company will confirm a word of it on the record.
Three months ago OpenRouter’s Series B set a reported $1.3 billion valuation. The two companies had already wired their products together back in April.
In today’s indie hacker news:
- 💰 Stripe clinches OpenRouter for more than $7 billion
- 📜 Two years of Anthropic system prompts, diffed
- 🦊 Firefox for iOS blocks ads, except search ads
- 🔌 Four tiers of RISC-V silicon under $100
- 🎚️ Your agent got dumber because a default moved
TOP STORIES
💰 THE ROUTER JUST GOT ROUTED
Stripe has reportedly clinched a deal to buy OpenRouter, the AI model router.
The story: Bloomberg reported Sunday that Stripe finalized an agreement for OpenRouter at more than $7 billion. OpenRouter sits between apps and model providers, handling routing, failover and cost optimization. A Stripe spokesperson said the firm does not comment on rumors or speculation, and OpenRouter declined to comment.
The price is the strange part. TechCrunch dates those talks, at about $10 billion, to the previous month. Fortune adds that the final number could still move. None of the coverage explains the drop.
The details:
- OpenRouter raised a $113 million Series B on May 28, led by Alphabet’s CapitalG. NVIDIA, Snowflake and Databricks venture arms all joined.
- Weekly volume climbed from 5 trillion tokens to 25 trillion over the six months to that round, per the company.
- The plumbing shipped first. Since April 29, one command,
stripe projects add openrouter/api, creates the account, the key and the Stripe billing. - Routing across 400 or more models is the product, plus failover for when a chosen model dies.
- CEO Alex Atallah co founded OpenSea, which raised more than $400 million and then watched usage collapse.
Why builders care: One vendor now owns both ends of your economics, the inference meter and the revenue. The buyer paid for the metering layer and not for a model, which says where the durable margin sits.
Nothing in the coverage says whether a router owned by a payments company still routes everywhere.
📜 READ THE DELETES, NOT THE ADDS
Two years of Anthropic’s shipped system prompts sit in public, and the diff is the lesson.
The story: Anthropic publishes the system prompts behind claude.ai and its mobile apps. The page carries 17 model entries, spanning July 12, 2024 to the Claude Opus 5 entry dated July 24, 2026. It hit the Hacker News front page with 573 points and 239 comments in about 15 hours. The submitter counted a bit more than 300 words in the earliest prompts and 3000 plus now.
Simon Willison has been diffing each release. Willison’s complaint: the page leaves out the tool definitions, and no Claude Code prompts appear at all. Willison reconstructed 23 named tools by asking Claude to list them.
“Anthropic are the only major AI lab to publish the system prompts for their user-facing chat systems.” (Simon Willison)
The details:
- The additions read as a failure ledger. Opus 4.7 gained a child safety block and a disordered eating section. It also got a rule against claiming a missing capability before calling tool_search.
- The deletions are the rarer signal. In the same diff, out went the line banning genuinely and honestly, plus a block naming the current president.
- The Opus 5 prompt hardcodes events that postdate training. One is the June 2026 export control suspension of Fable 5 and Mythos 5.
- Length is contested. One commenter says Pi’s prompt runs about 300 words. They claim it still beat Claude Code and Codex on token use and pass rate.
- A community archive at 63k stars carries what Anthropic will not ship. That includes Claude Code, Claude in Chrome and Claude Cowork captures.
Why builders care: Each addition marks a thing that broke, so the archive is a catalog of agent failure modes. The deletions are the rarer teacher, showing which rules a model outgrows once its training catches up.
That repo’s README classifies Anthropic’s published versions as the shorter ones, so read the page as a trimmed view.
🦊 BLOCKS ADS, EXCEPT THE PAYING ONES
Firefox for iOS is rolling out a built in ad blocker that waves through search ads.
The story: Mozilla is gradually rolling out an experimental ad blocker inside Firefox for iOS. AppleInsider reports it runs an EasyList filter and blocks at the network level. Third party ad networks, ad trackers, pop ups and overlays all get caught. Three taps in the site menu turn it on.
Two categories survive. Ads on search results pages at Google, Bing and DuckDuckGo stay, and so do sponsored tiles on Firefox Home. AppleInsider sets that beside Mozilla’s own disclosure. In 2025, 90% of its revenue came from Firefox, and roughly 85% of that came from a Google deal.
“So block ads unless it impacts their revenue sources.” (ImJamal, Hacker News commenter)
The details:
- The request is old. GitHub issue #13910 asked for exactly this on April 10, 2023, and it is still marked Open.
- The thread pulled 581 points, and most of the heat went to the carve outs rather than the feature.
- Mozilla telegraphed it in a June 2026 Reddit AMA, promising it was coming very soon with no date attached.
- Firefox Focus, Mozilla’s other iOS browser, already ships a system wide content blocker. One commenter dates that back to the late 2010s.
- Commenters kept recommending paid indie apps instead. Wipr came up most, at $5 a year for system wide blocking, alongside AdGuard and wBlock.
Why builders care: When a platform ships your feature for free, the refusals are the real product spec. Mozilla’s carve outs are tied to the money, so that gap stays sellable for as long as the deal holds.
Apple’s engine rule binds every iOS browser equally, so the edge is knowing the Content Blocker plumbing.
🔌 ELEGANCE DOESN’T SHIP, PRICE DOES
A Trinidad based embedded engineer answers the viral RISC-V critique with a receipt.
The story: Armstrong Subero, an embedded engineer in Trinidad and Tobago, answered Dmitry Grinberg’s RISC-V critique with prices. Subero covered four tiers of RISC-V silicon for under $100 including shipping, inside a year. One Segger J-Link debug probe on the ARM side runs about $600.
Grinberg’s original argues a high end CPU and a cheap microcontroller core want opposite things. So no single instruction set can serve both. It also concedes RISC-V will own the cheap microcontroller space anyway. The Lobsters thread argued more about tone than substance.
“Instruction set elegance is a thing you can afford to care about once the hardware is already on your desk. Whether the hardware can get to your desk at all comes first.” (Armstrong Subero)
The details:
- The bottom tier runs about $7 for 50 CH32V003s plus a debugger. That is a ten cent part with no multiplier and no divider.
- The middle tier is a $20 CH32H417 board with two cores at 400 MHz and 144 MHz. Subero reports running facial recognition on it in under 150KB of RAM.
- Up top sit a $9.50 Baochip Dabao and a $30 Orange Pi RV2. The Dabao runs a VexRISC-V with an MMU and was inside the DEF CON 34 badge.
- The ARM comparison is the point. The nearest equivalent, the STM32H747, carries a twenty week lead time and an ID check at Mouser.
- Shipping is the binding constraint. Landing one dollar chips in Trinidad costs $60 to $200, on parts that ship free elsewhere.
Why builders care: One base instruction set spans all four tiers, so the toolchain and calling convention carry upward. A solo hardware builder can prototype on the cheapest part and scale up without learning a new privilege model.
On ARM, virtual memory means licensing a different core family. On RISC-V, supervisor mode is an optional item somebody just added.
🎚️ TURN THE DIAL BACK UP
Your coding agent got worse last spring because a default moved, not because the weights changed.
The story: Walter van der Giessen argues that labs are trading world knowledge out of the weights, on purpose. The trade buys reasoning skill, and the evidence is score per active parameter. The post reports GLM-5.2 hitting 99.2% on AIME 2026 with about 40 billion active parameters. GPT-4 was rumored to run around 280 billion active in 2023, and could barely finish an AIME problem.
The product layer version already has receipts. Anthropic’s April postmortem says it flipped Claude Code’s default reasoning effort from high to medium on March 4. It reverted on April 7 and called that the wrong tradeoff.
“In general, the longer the model thinks, the better the output.” (Anthropic, April 23 postmortem)
The details:
- One system prompt line capped final responses at 100 words. It shipped April 16, cost 3% on one evaluation, and was pulled April 20.
- A caching bug reapplied a thinking clearing header every turn instead of once. Users saw forgetfulness and odd tool choices, and root causing took over a week.
- OpenAI ships the same lever as reasoning.effort with a medium default. Its docs tell developers not to treat it as the fix for bad output.
- On the weights side, the post says SimpleQA tops out at 53% with no tools. The best paid recall still misses about half the questions.
- Strip the stored facts and total size falls toward active size. The post puts a 20B to 40B model on the 24GB card already in gaming PCs.
Why builders care: The dial setting how smart your agent is sits in the product layer, not the weights. If labs keep shipping fact thin models, the moat moves to whoever builds the retrieval harness around them.
No lab in the sources says it is stripping knowledge on purpose, so the deliberate part is the author’s read.
TRENDING TODAY
- 🔍 The federal keyword lists that canceled billions in research funding - Higher Ed Dive got inside those lists. It pulled 143 points and 41 comments. The argument is about what happens when a grep decides policy.
- 🐢 Qwen 3.8 27B is excellent, but it defaults to overthinking things - Simon Willison rates the new Qwen highly. The catch is that it burns tokens reasoning when it does not need to. It drew 149 points and 61 comments overnight. A model that overthinks locally still beats one you rent.
- 📉 Nvidia reduces the OpenAI infrastructure financing it may guarantee - Reuters relays a Journal report on the pullback. Nvidia scaled back its $250 billion OpenAI data center guarantee. That drew 145 points and 57 comments. When the chip vendor stops underwriting its biggest customer, believe the balance sheet.
FIRST DOLLAR
💵 THREE CUSTOMERS, ONE DASHBOARD
$340 in MRR from three paying customers, three months in.
A Korean indie hacker published the whole journal from $0 MRR to $340 across three paying customers. The product is a KPI dashboard. It came out of freelancing for Korean startups where metrics lived in Notion, Excel and sticky notes. Validation was two weeks and 14 founder conversations, and the writeup keeps the mistakes in.
STACK OF THE DAY
🎙️ VocalCode
VocalCode is push to talk dictation aimed at AI coding agents, and it runs on device. Speaking a prompt beats typing one, and agent prompts run long. On device means the audio never leaves your machine, which matters if you dictate client details. It landed on Show HN overnight with barely any votes, so you are early on this one.
Not sponsored. We just feature tools builders would actually use.
BOOKMARKED TODAY
- 🎟️ The AI Credit Resale Economy - A field guide to the token brokers, meaning the people quietly reselling AI credits. It pulled 245 points and 94 comments. Read it right after the OpenRouter story, because it is the same business from the reseller’s side.
- 🧾 The Life and Death of Direct File - A Berkeley report autopsies the tool, from build through shutdown. It carries 112 points and 45 comments. Anyone who has shipped inside a large organization will recognize the shape.
- 🗂️ A local viewer for your ChatGPT and Claude exports - Someone got tired of not reading their own chat backups. So they built a browser for them, running fully local. The maker notes neither export carries enough to rebuild what the app shows. Projects are the biggest gap.
That’s the board for today. Go build something.
Ship to one country, QA it from another. Geo-fenced Stripe checkouts, region-gated APIs, feature flags that only fire in EU. NordVPN's 6,400+ exit nodes across 110+ countries let you test a US-only paywall from EU or hit a EU-only checkout from SF. Bonus: free Meshnet for SSH across your devices.
We get a cut if you sign up. Only added for tools we use ourselves.
Curated by AI, built by a human.