A speaker on the “Unprecedented Scale” panel at AI House Davos 2026 said something that stuck with me: teams are building with their tools, on top of open AI and others, working alongside them, and moving very fast. Then the kicker — “I really believe in the ambition of European founders.”
Ambition is lovely. Ambition does not pay inference bills. And that gap between what founders say on stage and what they actually ship is the whole story of the open-versus-closed argument that played out at TechCrunch Disrupt 2026, where founders spent real stage time debating which side of the fence to build on.
The benchmark argument is over, and that matters less than people think
Here’s what’s genuinely changed. As of early 2026, open-source AI hit what Towards AI calls a real inflection point. The gap between closed and open models didn’t just narrow — on many benchmarks it closed entirely. That’s not a vibes claim. That’s the measurable part.
So if you’re a founder choosing a foundation model and your decision criteria is “which one scores higher,” congratulations, you’ve been freed from a decision that no longer exists. Pick either. The benchmark tiebreaker is dead.
Which is exactly why the Disrupt debate got interesting instead of boring. Once capability stops being the differentiator, founders have to argue about the stuff they were previously allowed to ignore: cost structure, data control, vendor dependency, deployment constraints, and whether you want your margins set by someone else’s pricing page.
I’ve reviewed enough agent products to tell you which of those actually kills startups. It’s rarely model quality. It’s the pricing page.
Watch the conflation that’s already spreading
One thing I want to flag before it becomes received wisdom, because I’m already seeing it bounce around.
Anthropic published figures showing Claude improving on open-ended problems — from roughly 26 percent to 91 percent, with the sharpest jump in March 2026 after internal access to Mythos Preview, landing somewhere around 88 to 92 percent by September 2026. Those are strong numbers on a hard category of task.
They are also not numbers about open-source anything. “Open-ended” means the problem has no single correct answer. “Open-source” means you can download the weights. Different words, different meanings, and I’ve watched at least one person on my timeline cite the first as evidence for the second.
If you’re a founder building a thesis, build it on the actual claim. Open models closed the benchmark gap. A closed model got much better at fuzzy problems. Both true. Neither proves the other.
Foundation Capital’s architecture bet is the more useful signal
Buried under the open-closed shouting is a prediction I think founders should take more seriously than the licensing debate. Foundation Capital argues that in 2026, “Cursor for X” becomes the default architecture for knowledge work — legal, finance, marketing, sales, operations. Not chat interfaces. Something that blends open-ended exploration with constrained precision, and won’t feel like a chatbot.
Read that carefully, because it quietly reframes the whole question. If the winning shape of an AI product mixes loose exploration with tight, verifiable precision, then you’re not choosing one model. You’re choosing a stack. Exploration and precision have different failure modes, different latency tolerances, and different cost profiles.
Which makes “open or closed” a weirdly flat way to frame your decision. The honest answer for most products shipping this year is both, in different places, for different reasons. The part of your product that needs to roam gets one treatment. The part that needs to be right, every time, auditable, gets another.
Founders who pick a side as an identity are going to lose to founders who pick per component.
What I’d actually tell you
My read, and you’re free to disagree:
- Stop treating model choice as a philosophical position. It’s a procurement decision with a technical attachment.
- If benchmarks have converged, your differentiation moved somewhere else entirely — product surface, workflow fit, evaluation discipline. Spend your energy there.
- Assume you’ll swap models at least once. Build the abstraction layer early, however unglamorous that feels.
- Be suspicious of any open-source argument that leans on numbers from a closed model’s progress report.
The Disrupt stage rewards conviction, and conviction sounds like picking a side. I’d rather see founders admit the less quotable truth, which is that a solid 2026 product probably runs open weights where control and cost matter, calls a hosted API where raw capability matters, and keeps the option to flip either one when the pricing changes.
That doesn’t make a good panel. It makes a company that’s still here in 2027.
🕒 Published: