Your Best Feature Is Someone Else’s Sprint

TypeSafe AI launched Jev on September 15. It is a model built to return structured decisions rather than paragraphs, and by most accounts it was picked up faster than any model in AI Gateway history. The team is led by Diogo Almeida, a former OpenAI researcher who worked on ChatGPT. Good founder, real product, obvious demand.

Six days later, the most widely read piece of writing about Jev was a blog post asking whether OpenAI would eat its lunch. It hit the Hacker News front page with 271 points and close to 200 comments. The argument took about two paragraphs to make. OpenAI has quietly run its models as classifiers for years. It has simply never sold classification as a standalone product.

That is fast-follow risk, stated more plainly than anyone will ever state it to you. And if a team with that founder and that adoption curve gets the question asked inside a week, your seed-stage company is going to get it asked too. Probably in a first meeting. Probably by someone who has not used your product.

Fast-follow risk is the chance a larger company ships something close enough to your product, quickly enough, using assets it already owns.

Three qualifiers there, and each one matters. Close enough, not identical. Quickly enough, not instantly. Assets it already owns, which is the part first-time founders consistently underrate.

The incumbent does not need to beat you. It needs to shrink the gap between your product and the thing your customer already pays for until switching stops feeling worth the paperwork. Most startups do not lose to something better. They lose to something adequate that was already installed.

So the real question is never "could they build this?" Of course they could. The question is whether they can build it with people, data and distribution they have on hand today.

Jev is a real product category rather than a wrapper. Decision models, tuned for judgment-heavy calls inside software, are a different shape from chat. The launch coverage in The Rundown and elsewhere treated it as a new class of model, not a repackaging.

And yet. The case against it is that OpenAI has the same capability sitting unshipped inside its existing stack. Not a research gap. A packaging gap. Packaging gaps close in weeks.

Look at what else happened that week. On Monday September 22, OpenAI shipped GPT-6 Sol and Luna. Anthropic shipped Claude Opus 5.5 within hours of it. Both posts sat at the top of Hacker News with more than 1,200 points each. Two frontier launches inside a single day.

The cadence is the threat, not any one launch. When the platform underneath you ships every few weeks, the surface area it can absorb grows every few weeks too.

Four properties, and you can check all four this morning.

Your product is a packaging decision rather than a discovery. If the hard part was noticing that a capability should be sold on its own, somebody else can notice it after you do.

Your users already hold an account with the incumbent. Nobody has to be convinced to sign up. They have to be convinced to click one more toggle.

Your data is public, purchased, or generated by the incumbent anyway. If you scraped it, they can scrape it. If you licensed it, they can write a bigger cheque.

Your switching cost is one base URL. This is the brutal one for anything API shaped. A migration that takes an afternoon is not a moat, it is a speed bump.

Score yourself on those four without flattering anyone. Three or four yes answers and your defensibility is currently a head start, which is a real asset but has an expiry date on it.

Not the model. Almost never the model.

Data that only exists because people use you. Corrections, labels, outcomes, edge cases your customers fed back in. An incumbent can buy a dataset. It cannot buy the record of what your specific users decided last Tuesday and whether it worked.

Workflow depth. If you replaced a call, you are replaceable. If you replaced a process with five approvals, two integrations and a monthly report somebody presents to a board, you are load-bearing.

Distribution the incumbent will not staff. Regional resellers, a trade association, an implementation partner network, a niche conference nobody at a large lab would attend. Distribution is boring and that is exactly why it lasts.

Contractual and regulatory position. Certifications, data residency, procurement approvals, a signed three-year term. Slow to earn, slow to lose.

A name inside one narrow segment. Being the obvious choice for veterinary clinics beats being the eighth-best general tool.

None of those show up in a demo. All of them show up in renewal rates.

Write the incumbent's build spec for your product. One page. Pretend you are a product manager there and you have been told to ship something competitive by the end of the quarter.

Be fair to them. List the team they already have, the data they already hold, the customers they can email tomorrow. Then write the ship date you would commit to.

If that date is inside a quarter and the team is people they already employ, you are exposed. That is not a reason to quit. It is a number to plan around.

Now write the second list, which is the one that matters. On the day they ship, what are they still missing? Every item on that list is your actual company. Everything else is a feature you happen to ship first.

Keep both lists somewhere you will look at them again. A doc works, a spreadsheet works, and a planning tool like Foundra gives first-time founders a structured place to keep competitive work sitting next to the financial model and the go-to-market plan instead of rotting in a folder.

Five moves, roughly in order of how often they work.

Go deeper into the workflow. Stop selling the capability and start selling the outcome around it, including the boring parts nobody wants to build. Reporting, approvals, audit trails, the export that finance needs.

Narrow the segment until you are unmistakably the specialist. A large company will not build a version for 4,000 orthodontists. You can.

Move to the data layer. Instrument outcomes so that after nine months of use, your customer's own history is the thing they cannot take with them.

Sell to the incumbent's enemies. Every platform has a set of customers who will pay a premium not to deepen that relationship. That is a real market and it is usually underserved.

Change the shape of the business. Services-heavy, outcome-priced, or bundled with something physical. Harder to copy because it is unpleasant to copy.

Doing nothing is also a choice. It is the one most teams accidentally make.

Sometimes it is fine. Dropbox shipped into a category every large platform could obviously serve, and did well for years because the category was enormous and the execution gap was wide. Slack did the same.

The test is not whether you can be copied. It is whether the market is big enough that a copied version still leaves you a business, and whether you can reach durability before parity arrives.

So ask two things. How long is your head start in months, measured by how quickly the incumbent has moved historically? And what do you need to have built by the time it ends? If the real answer to the second question is "more users," you are in trouble. If it is "a dataset, 40 signed annual contracts, and a partner channel," that is a plan.

Plenty of good companies were built on a two-year window. Very few were built on a two-year window nobody had measured.

• Fast-follow risk is about assets the incumbent already owns, not about whether the work is hard.
• A one-afternoon test: write the incumbent's build spec, then list what they would still be missing on launch day.
• Defensibility usually sits in usage data, workflow depth, distribution and contracts. Rarely in the technology.
• An API-shaped product with a one-day migration path has a head start, not a moat.
• Two frontier model launches landed on the same Monday. Assume the platform under you moves at that speed when you plan a year.
• If the exposure answer is bad, pick a move now. Narrowing the segment is the fastest one available to a small team.

Should I avoid building anything a big lab could ship?

No, because that rules out almost everything worth doing. Build it, and be clear-eyed about what you need to own by the time they arrive.

Investors keep asking what happens if OpenAI builds this. What is a good answer?

A specific one. Name what they would ship, name the date you think they would ship it, then name the three things you will have that the shipped version will not. Vague confidence reads worse than acknowledged risk.

Is having a former researcher from a big lab on the team a defence?

It helps with hiring and with raising. It does not stop a fast-follow. Jev has exactly that and got the question anyway.

How long is a typical head start in this market?

Short, and shortening. Plan against months rather than years unless you have evidence to the contrary in your own category.

Does a patent help?

Occasionally, in hardware and life sciences. For software workflows it is usually slow, expensive, and less useful than a signed multi-year contract.

What is the single best early signal that I have real defensibility?

Customers who renew after a cheaper adequate alternative shows up, and can tell you in their own words why they stayed.

1. Will OpenAI Eat Jev’s Lunch? Arcturus Labs, September 21, 2026
2. OpenAI is well positioned to fast-follow Jev, Hacker News discussion, September 22, 2026
3. Ex-OpenAI Researcher Launches Jev, A Chatbot That Makes Decisions Instead Of Text, ETV Bharat
4. TypeSafe launches Jev for AI decisions inside software, The Rundown AI
5. Introducing GPT-6 Sol and Luna, OpenAI, September 22, 2026
6. Claude Opus 5.5, Anthropic, September 22, 2026