“No Vendor Lock-In” Is Code for “No Product” — Ferran Sulaiman

“What occurs when Claude is not essentially the most clever mannequin?”

“What occurs if Muse simply will get actually dangerous tomorrow?”

“What if OpenAI simply jacks up the costs of those fashions?”

These are all arguments I’ve heard for when individuals say that they are constructing an present product with a mannequin picker.

And to that I say “bull.”

“No vendor lock-in” is the weakest reply you could give when your product is a clone of a profitable product with a mannequin picker.

I’ll inform you why.

Today any product that employs AI (or AI brokers) to do something for a particular use case is a customized harness. @ycombinator mentioned as a lot not too long ago in one in all their Paper Club talks not too long ago. And it additionally appears to be like like a lot of the newest YC batch is constructing a customized harness for one explicit use case. And I agree. I feel that is the correct technique to go as properly.

However what’s new is a complete suite of merchandise whose complete pitch is that you are able to do the identical factor that you are able to do with one of many merchandise from these frontier labs “however you’ll be able to select from our listing of 420,69 fashions that we provide”. Think Claude Code, however with different fashions. Think ChatGPT working with different fashions. Think Grok bot with different fashions.

And the one declare for the existence of such merchandise is that in the future Anthropic or OpenAI will jack the costs up till the mannequin is unaffordable or the mannequin will simply grow to be unusable.

The different declare is that open-source fashions exist and that they are low-cost and therefore we needs to be utilizing them.

The drawback with all these claims is that they assume that we use frontier fashions as a result of we belief the businesses or as a result of the value is true.

That’s not why.

We use them as a result of we wish the perfect mannequin out there. We need GPT-6 Astra, Claude Fable 5.1, and Opus 5.5. They’re actually good fashions and open-source fashions aren’t even shut. We solely speak about these open-source fashions across the time they’re launched they usually slowly fall off as a result of the frontier fashions are simply so good and accessible.

I used to be in opposition to Claude Fable.

I believed it was too costly.

I believed Kimi K3 might do all the things.

Then I did get round to utilizing Fable. I came upon that Fable is the perfect options engineer that I might ever ask for and nothing else I’ve used is even shut and that is telling as a result of I’ve used virtually all of them. Astra is a very actually good coworker and nothing else has come shut in that sense both.

Benchmarks say the identical factor: open sources are closing the hole however they’re nonetheless nowhere close to the frontier fashions as they exist right now.

Building a product on the weaker fashions and pitching the mannequin listing is the worst wager you would make. Everybody simply needs the perfect efficiency. Nobody cares concerning the mannequin tier listing beneath.

People ought to have entry to the perfect efficiency always and it means having access to the very best mannequin for the duty always. They shouldn’t have to change to a less expensive mannequin to get the job accomplished.

Cheaper right now nonetheless signifies that the output is (typically, barely) worse. If it wasn’t, open supply would already be the default.

We used to guage fashions on price per million tokens. We now decide them on price per activity and that is the quantity I belief. A decrease price per activity wins even when the associated fee per million tokens is a bit increased, model-wise.

What issues is the product. What’s the correct UX for activity X? What has my consumer been caught on each time they attempt to do operation Y? How will we get them previous that?

The one product that is completely killing this right now, for my part, is Codex or ChatGPT work. You nonetheless know which mannequin it makes use of beneath however actually the common consumer is healthier off not realizing as a result of it conforms to the correct consumer expertise for no matter you ask. If you need to generate a picture, it pops up this stunning picture expertise the place you’ll be able to annotate and inform your agent what it wants and what it must do. If it is a doc that you just need to edit, now it opens up in OpenAI areas, and so forth.

If OpenAI weren’t a mannequin lab they would not actually even give the consumer the choice of a mannequin picker. The job is handy the consumer the end result. They do this regardless of which mannequin is chosen beneath.

Anthropic is shifting the identical approach. Their laptop use has gotten actually good. Companies successful exterior the lab are doing the identical factor. I like T3 Code. It makes software program engineering pleasurable. I plan in it. I architect in it. The brokers write the code.

Paper received the product expertise proper as properly. I’m a software program engineer and with Paper I can design on a canvas, use no matter agent I’ve, and share it. Most of the time I do not care which mannequin it’s. It works inside Codex, inside Claude Code, or some other harness of my selection.

I used to trace each Apple launch and mainly each tech product between 2010 and 2021 however then I ended as a result of most of them received ok that purchasing any of them right now could be advantageous. I can nonetheless get the job accomplished. I solely listen when one thing is definitely new. Models are heading the identical approach. Which mannequin you utilize issues lower than what you construct with it.

Build a product individuals love to make use of. Make the perfect product win.

I’m additionally not saying we should not use open-source fashions. I like them. They’re low-cost they usually work. Use them. All I’m saying is it would not matter which one you find yourself utilizing, so long as the job will get accomplished.

The consequence is way extra necessary for our customers than letting them select which mannequin they need to use.