Models got cheap. Proof just got expensive.

The week of July 27, 2026

The signal (TL;DR): In July, Cursor rebuilt SQLite from an 835-page manual with agent swarms. Every passing configuration shipped the same working software, but the bill ranged from $1,339 to $10,565 depending on which models did the work. Days later, word broke that Stripe is in talks to buy OpenRouter, a model-routing layer valued at $1.3B in May, for roughly $10B. When any model can do the job and a router picks one mid-request, the scarce thing stops being intelligence and becomes proof of what ran. This week’s play: sell a five-day Model Receipt Packet to one small vertical SaaS vendor staring down a customer security review. Pilot price: $350. The free Model Receipt Kit is below.

Moonshot’s Kimi K3 released its full weights on July 26, a 2.8-trillion-parameter model anyone can download. More than half the tokens flowing through OpenRouter-style trackers already go to open models, most of them Chinese. And as of July 24, Polymarket priced a 45% chance the US bans an open-source model this year, up from 22% days earlier.

Read those three lines together. The engine behind AI software is now cheap and swappable. It just went politically unstable too.

Every other newsletter will tell you what that saves you. This issue is about what it makes you: the person who can prove which engine ran, because almost nobody can.

The play pays $350 for a week of careful work, and it compounds. Let’s get into it.

Start with the experiment.

In July, Cursor set agent swarms loose on a brutal task: rebuild SQLite, one of the most-used pieces of software on earth, from its 835-page manual. Multiple configurations passed every test. Same outcome, same quality bar.

One configuration cost $10,565. Another cost $1,339.

An 8x price spread for identical output, decided entirely by which models were arranged in what order. That spread is why routers exploded the week of July 20: Ramp opened the router it ran internally for three years, Cursor shipped one claiming frontier-level results at 60% lower cost, and Meta is building its own.

The engine went generic. On All-In on July 24, Chamath Palihapitiya said he runs the same jobs through Grok, GLM, Claude, and others and “can’t even tell the difference.” The panel’s rough consensus: something like 95% of tasks can be done by many different models. Microsoft’s numbers agree. Its small harness-trained models beat frontier models per dollar, and swapping one image model cut PowerPoint’s image costs 84%, per Mustafa Suleyman in July.

Then the meter arrived. Uber capped employee AI spend at $1,500 a month back in June. Tesla runs $200 a week with a request form for more. Ryan Carson, running a one-person legal-tech company on AI agents, told The Startup Ideas Podcast on July 24 that he burned about $20,000 in tokens last month and expects the market to settle near $5,000 per employee. The token-max era is ending in budgets and caps.

And Washington made the model supply chain political. The White House says K3 was distilled from Anthropic’s Fable. Treasury floated sanctions. Nearly 200 startups signed a letter warning that a ban on Chinese models would devastate the startups built on them. Whatever side wins, one fact is already priced in: the model behind your software can be repriced or pulled entirely, and your customers never see a change on screen.

So picture the stack under any AI feature in 2026. A router. Several eligible models, some open, some closed, some foreign. Fallbacks. Policies that change weekly.

Now picture the enterprise buyer’s security questionnaire landing on that stack. Who processed our data? Which models were eligible? Were our inputs retained or trained on? Can you prove which model handled a given request?

For most small software vendors, the honest answer is half router settings and policy pages, half one engineer’s memory. Cheap intelligence created that gap. It did not create anyone to close it.

Effort: 3/5 · Cost to start: $0-30 · Time to first $: 3-10 days · Skill: careful research, structured interviews, clear writing. No code.

Buyer. A 5-30 person vertical SaaS company that isn’t an AI company but added one AI feature: summarizing, classifying, drafting, search. Field-service software. Property management. Construction back office. The tell that they can pay: they sell an enterprise plan, and a real customer security or procurement review is sitting in their inbox.

Pain. The founder can explain what the feature does. Nobody can cleanly answer what runs under it. David Pan of Cursor described the era his own customers just lived through, in July:

That insanity is over for choosing models. It never got solved for documenting them. An overconfident answer on a security questionnaire is worse than a slow one: it creates a claim the vendor can’t defend later.

Offer. In five business days, you turn one live AI feature into a Model Receipt Packet: a one-page data-path map, a register of every eligible model and provider with dated policy evidence, the route-receipt fields their logs actually capture, a plain-language buyer FAQ, and a change card naming who reviews what when a model gets swapped. Every claim links to dated evidence. Every gap is labeled unknown instead of papered over.

You are selling a faster, defensible first response to a real buyer review. You are not selling compliance, a legal opinion, or a security certification. Say so in writing.

Who pays whom. The SaaS company pays you $350 for the pilot: $175 before the technical interview, $175 at delivery. After two accepted pilots, the standard sprint is $700, plus an optional $125 quarterly refresh when policies or routes change. Your costs are $0-30 in document tools. Modest case: one pilot, 8-10 hours of work, at least $320 gross before your time and taxes. Don’t model a retainer until the first packet survives a real review. I think $350 is underpriced for this, and that’s deliberate: the pilot’s job is proof, and proof reprices you.

Why now. The routing layer just became the most fought-over real estate in AI. Stripe’s reported ~$10B bid for OpenRouter values a two-month-old $1.3B round at nearly 8x, because whoever owns routing owns the record of what ran. That record is exactly what your client’s buyer wants, and almost no small vendor can produce it today.

The proof. The packet’s credibility comes from its honesty. If the client’s logs don’t record which model served a request, the packet says so, marks it UNVERIFIED, and assigns an engineer to fix or narrow the claim. A guessed assurance is worthless. A dated, owned gap is something a procurement team can work with.

A worked example. FieldNote, a hypothetical 12-person field-service SaaS, added an AI note summarizer. Its router can send requests to either of two models. Its logs capture cost and latency but not which model answered. The packet’s FAQ entry reads: “Can FieldNote identify which model processed a specific request? Not from the current application log. The router configuration shows two eligible routes, but the selected route is not preserved in the log sample provided. Status: unverified. Owner: engineering. Next decision: preserve the selected provider/model per request, or narrow the claim made to customers.” That answer is less polished than a promise and far more defensible. Ten answers like it, and the review moves.

First move (48 hours). Build the FieldNote specimen yourself from the free kit below, clearly labeled hypothetical. Then list 20 small vertical SaaS companies that announced an AI feature and sell an enterprise plan. Message the founder: “A routed AI feature can change providers without changing the button your customer clicks. When a buyer asks who handled their data, the answer usually lives across product settings, vendor policies, and an engineer’s memory. I turn one live feature into a source-backed Model Receipt Packet for the review already on your desk. Pilot is $350. If no customer is asking, you shouldn’t buy it. Here’s a specimen of exactly what I deliver.” Ask to see the live questionnaire before taking the deposit.

The honest part. The receipts above prove routing, caps, and policy risk are real. They do not prove small SaaS vendors already pay beginners for receipt packets. That’s the bet, and the pilot is the test. You’re not a lawyer or an auditor, and the packet never touches credentials or production data: the client’s engineer supplies screenshots, config exports, and policy links. If 20 qualified contacts with a live enterprise motion produce fewer than three conversations and zero paid pilots, kill it and keep the skill.

The Model Receipt Kit builds the whole packet with you, specimen first (grab it free). Inside:

Do it in 5 minutes. Open the kit, copy the FieldNote specimen, and swap in a fake company from a niche you know. That one artifact is your portfolio and your practice run.

The specimen is the pitch. The packet is the product.

If you already do client work. Add the packet as a $350 wedge on top of whatever you sell. It’s a five-day deliverable that puts you inside the client’s AI stack, and every model swap after that is a reason they call you back.

If you’re starting from zero. This play needs no code and no reputation. The FieldNote specimen does the credibility work: you’re showing finished output, not a resume. One accepted pilot beats any portfolio site.

If you run the AI budget at your company. The second payroll line is forming whether you plan it or not. Budget it like one, and demand the receipt fields from your own vendors before your buyers demand them from you.

If you sell software with an AI feature. Build your own packet with the kit before anyone asks. The vendor who answers a security review in two days instead of three weeks wins deals on speed alone.

This play lives in a window. The routers can see everything the packet documents, and I put the odds that a major router ships a native customer-facing provenance report within a year at better than even. That call goes on the scorecard. Until it ships, and for every vendor whose stack spans more than one router, the packet is manual work someone must sell.

The window is the play. Documentation gaps close; the operators who closed them keep the clients.

The bear case is bigger than a feature release. Steve Keen, the economist known for calling the 2008 crisis, told the PBD Podcast on July 21 where he thinks all of this is headed:

If he’s right, AI budgets shrink and procurement reviews get meaner: fewer AI features get bought, but every one that survives gets audited harder. The packet is one of the few AI services that gets more relevant in a downturn. What doesn’t survive a downturn is vague spend. Don’t build your income on token-maxing clients.

And this is not for everyone. If reading a retention policy bores you into skimming, skip this one: the product is care, and skimmed care is a liability.

The week of July 20 drew one line: the model layer went generic (an 8x price spread for identical software), the meter arrived ($1,500 caps to $5,000-per-employee budgets), and the plumbing that decides what runs is being bought for ~$10B. Abundance in the engine room creates scarcity at the front desk. Somebody has to prove what ran, in writing, with dates. For about a week of careful work, that somebody can be you, at $350 a packet, in a niche where nobody else is even offering.

Freeze the scope. Date the evidence. Sell the receipt.

New plays land Monday mornings. Subscribe free.

The locked library is bigger than this week.

This week’s play is free. The other 45 in the Pro Vault are not, and a new one banks most weeks. Each is a complete play with the prompts, templates, and checklists to run it. The essay always stays free. The runnable kit retires into the Vault when the next issue ships.

The deeper version is the part I'd never post publicly.

Free gives you the play; Pro gives you the deeper build. This issue’s is The Proof Room: the done-for-you provider register with dated policy evidence for the major model vendors, the 40-question buyer-FAQ bank mined from real security questionnaires, the filled FieldNote packet as an editable template, and the outreach-to-retainer pricing ladder. It gets built out with the founding cohort, on top of the four finished builds already on the shelf: the Spread Desk, the Pack Line, the Operator’s Edition, and the Rep Desk. Plus the community that opens when the founding cohort is in, and the forward-only scorecard where misses stay on the board.

The founding price never changes.

Founding members lock $129 a year, forever. The founding rate closes when the first 25 members are in. After that, the standard price rises toward $399 as the library and the scorecard grow. The Vault opens the moment your license key lands.

Checkout takes a minute. Your license key arrives by email and opens the Vault.

Out-yield the average.
Javier @ Overyield

Know a founder who just bolted AI onto their product? Forward this. Their next security review will thank you.

Overyield is educational, not financial, legal, or business advice.