Stripe is in talks to acquire OpenRouter (the model-routing layer) for close to $10B, up from a $1.3B valuation; OpenRouter also held early acquisition talks with Databricks. Reported by Stephanie Palazzolo and Amir Efrati (The Information). Update to our July 17 "OpenRouter sale for billions" item; the delta is the buyer (Stripe) and the number. https://x.com/steph_palazzolo/status/2080439223115538596 | https://www.axios.com/2026/07/24/stripe-openrouter-merger-ai-currency
Read-through for your bill: the switch between models is now owned by a payments giant. Keep your OWN routing config so you never inherit one aggregator's pricing or terms.
A system, not a freelancer: a 30-year operator (real estate, hospitality, food, a theater) directs a full team of AI agents at one client project at a time, strategy to ship. The agents do the work, one senior operator steers the fleet, you own the code and the keys at the end. When the best model halves in price on a Saturday and the routing layer gets bought on a Thursday, you want a driver. Book a 30-min call. SPONSOR: https://outpace.media
The stories
LEAD (Claude, update), Opus 5, the 24-hour verdict + the prompting change
We covered the launch July 25 (half Fable's price, #1 intelligence, best computer use). DELTA today: a full day of real builds. Theo: "a strange but useful in-between of GPT-5.6 Sol and Fable 5... takes things super literally... code slightly worse to look at but more likely to be correct... regularly catching things Fable missed." Every's vibe check: brilliant in flashes, argues with instructions, doesn't always play nice with existing skills/plugins. Matt Shumer one-shot a full three.js game, all custom code, zero external assets.
The move: (1) rewrite fuzzy prompts into explicit specs, Opus 5 won't infer intent you left out; (2) re-test your skills / CLAUDE.md files against it before trusting them, some were tuned to Fable's guessing and now fight you; (3) lean on its self-checking, ask it to verify before it ships.
Codex multi-folder local projects (OpenAI, July 24-25)
A local project can now include related code, docs, and reference files from multiple folders; Codex reads/writes across all of them while one primary folder stays the Git root.
The move: add your reference/design/docs folders alongside your code root so the agent stops working half-blind.
Spend 10 minutes turning your most-repeated task into an explicit spec, then run it on Opus 5 and watch a literal model nail it. The spec outlasts the model.
The written brief from each episode, free, every morning.
Transcript
0:00Nova: Twenty-four hours with Opus five and the verdict is in from people who actually built with it, and here is the headline: it is not Fable. It stopped guessing what you meant, and that one change rewrites how you prompt. This is Mainframe, your daily guide through the AI chaos: what actually happened, who is winning, and how to take yourself to the next level. It is Sunday, July twenty-sixth, twenty twenty-six. I am Nova.
0:29Nova: On today's Mainframe: Opus five, one day in, the hands-on verdict from real builders and the single prompting change that makes it sing. Codex can now reach across folders, your code, your docs, your reference material, all one project. Your Python linter went from fifty-nine default rules to four hundred thirteen overnight, so today it is going to light up your repo, and we have got the batch-fix move. And a research-folder pattern from Vercel's Guillermo Rauch, plus a free way to run Opus five with no new bill. Let's get into it.
1:11Dex: One honest note before the board: these voices are AI. The reporting, the picks, and the calls you are about to hear are human-made. Now, the board.
1:21Nova: One macro move that touches your stack. Stephanie Palazzolo and Amir Efrati at The Information report Stripe is in talks to buy OpenRouter, the model-routing layer, for close to ten billion dollars, up from a one point three billion dollar valuation, with Databricks also in early talks. We flagged the OpenRouter sale as "billions" on the seventeenth; the delta is the buyer and the number. Read-through: the switch between models is now a payments-giant asset. Own your own routing config so you never inherit somebody else's pricing.
1:59Nova: It is. Brought to you by Outpace. Picture the person who spent thirty years building real businesses: real estate, hospitality, food, even a theater. Today they point a whole team of AI agents at one client's project at a time, strategy to ship. Outpace, at outpace dot media, is a system, not a freelancer and not a faceless agency: the agents do the work, one senior operator directs the fleet, and you own the code and the keys at the end. On a week when the best model on the board halves in price on a Saturday and the routing layer gets bought on a Thursday, you do not want to be steering that alone. So here is the move: book a thirty-minute call at outpace dot media. That is outpace dot media.
2:52Nova: Lead story, and it is an update, so only the delta. We covered the Opus five launch yesterday: half of Fable's price, narrowly number one on intelligence, best-in-class computer use. What is new today is twenty-four hours of real builds, and the verdict is sharper than the spec sheet. Theo put it best: Opus five is a strange but useful in-between of GPT five point six Sol and Fable five. It takes things super literally, its code is slightly worse to look at but more likely to be correct, and it regularly catches things Fable missed. Every's vibe check agrees: brilliant in flashes, and it will argue with your instructions and does not always play nice with your existing skills and plugins.
3:40Dex: So the model got more capable and less agreeable.
3:43Nova: Right, and that is the actionable part. Fable would read your mind and do the tasteful thing. Opus five does exactly what you wrote, no more. So the move this week is three steps: one, rewrite your fuzzy prompts into explicit specs, because Opus five will not infer the intent you left out. Two, re-test your skills and CLAUDE dot M D files against it before you trust them, because some were tuned to Fable's guessing and will now fight you. Three, lean into its self-checking; it catches its own errors, so ask it to verify before it ships. And the upside of literal-and-correct: Matt Shumer one-shot a full three J S game with Opus five, every asset custom-coded, nothing external. Detailed spec in, working game out.
4:40Nova: OpenAI's Codex team shipped multi-folder local projects. Until now a Codex project was basically one directory; now a project can include related code, docs, and reference files from several folders at once, and Codex reads and writes across all of them while one primary folder stays the Git root. If you have ever watched an agent lose the plot because your API docs lived one folder over from your code, this is the fix. The move: add your reference and design folders alongside your code root so the agent stops working half-blind.
5:18Dex: Third, the one that is going to surprise people when they open their editor Monday.
5:22Nova: Astral shipped Ruff zero point sixteen point zero, and the fast Python linter just raised its default rule set from fifty-nine rules to four hundred thirteen. Simon Willison ran it and it surfaced one thousand six hundred eighteen problems in one project alone. This is not a bug, it is seven times more checks on by default, so your clean repo is about to look messy. The move, and it is a good one: run it, then hand the whole report to your agent and have it batch-fix. You will clear months of latent issues in one pass.
5:57Dex: To the tool lab. Two picks, one technique, one thing.
6:00Nova: The technique comes from Guillermo Rauch at Vercel, and it is dead simple. He does all his research now with an agent CLI and the filesystem: a folder literally called research, an AGENTS dot M D file inside it describing the format he likes and his best practices, and then he launches the agent and asks a question. The agent finds and correlates knowledge from previous sessions, and it scales to infinity because it is just files. No app, no knowledge graph, no UI. Sync the folder over iCloud or git and it follows you across machines. Set it up today: make a research folder, write one AGENTS dot M D with your preferred output format, and point your CLI at it. His framing is that the factory is the product, and this is the smallest possible factory.
6:56Nova: This one is a hidden gem, and it is timely: T three Code, from Theo and the Ping team, open source and free. It is a clean graphical front end for Claude Code and Codex that runs on the subscription you already pay for, no extra bill, no key to buy. You install it with N P X T three at nightly. In the last week they merged seventy-six pull requests: Opus five support, an Auto mode for approvals, and Claude Code skills right in the composer. So if you have been living in the terminal, this gives you parallel agents, a real thread list, and worktree isolation on top of the exact subscription you own. First step: run it tonight against a repo you know and drive two agents at once.
7:47Dex: Use cases. Ways to use what you already have.
7:49Nova: Three. One, the Opus five game build, credit Matt Shumer: this weekend, hand Opus five one detailed spec for a small canvas or three J S toy, tell it no external assets, and let it generate everything. The literalism that makes it annoying in chat makes it a monster at one-shotting a tight spec. Two, the Ruff cleanup, credit Simon Willison: run Ruff zero point sixteen on your Python project, pipe the report into Claude Code, and have it fix the flagged rules in batches with tests between each batch. Three, consolidate your scattered repo: use the new Codex multi-folder support to pull your docs and reference code into one project so the agent finally sees the whole picture instead of one directory.
8:41Dex: Which is the whole game now, being findable.
8:44Nova: It is. Brought to you by SearchVis, also an Outpace Media product. Mainframe tells you what shipped; SearchVis makes sure the models know what you shipped. When a buyer asks Claude or ChatGPT or Perplexity for the best tool in your category, the model names you or it names your competitor, and if it does not name you, you do not exist to that buyer. SearchVis tracks whether the answer engines cite your brand across every engine, tells you why, and hands you the exact move to win the citation. Check where you stand and start free at searchvis dot outpace dot media. That is searchvis dot outpace dot media. Be the answer.
9:29Nova: Here is the through-line nobody framed out loud today. Opus five stopped guessing. Rauch says the factory is the product. Both are the same lesson: the model is done reading your mind, so the value moved to the precision of what you hand it. For two years the winning move was a clever prompt that nudged a model toward the tasteful answer. That era is closing. The new models are literal and correct, which means they reward an exact spec and punish a vague one. The scarce skill is not prompting anymore; it is specifying, writing down what "done" actually means so a literal machine can execute it without inventing the parts you left blank.
10:13Dex: So the fuzzy prompters are about to have a bad month.
10:16Nova: They are, and the people with a written spec are about to feel unstoppable. So here is what I would actually do: take the one task you keep re-explaining to your agent, and this week write it down once, as a spec or an AGENTS dot M D, precise enough that a literal model gets it right the first time. Stop reprompting. Start specifying. That file outlives every model on the leaderboard.
10:45Nova: One move this week: spend ten minutes writing your most-repeated task as an explicit spec, then run it on Opus five and watch a literal model nail it. And if you want every story, number, and link from today in your inbox, subscribe to the free Mainframe newsletter, the written brief, every morning.
11:03Dex: That is the show. We will see you tomorrow.