From 26 percent to 91 percent. That’s the jump Anthropic reported in Claude’s performance on open-ended problems, with the sharpest climb arriving in March 2026 after internal access to Mythos Preview. Hold that number next to another one from early 2026: on many benchmarks, the gap between closed and open models hasn’t just narrowed, it has closed entirely.
Two numbers, two directions. One says the frontier labs are still pulling away on the hardest class of problem. The other says the open models have caught up on the measurable stuff. If you’re a founder trying to decide what your product sits on top of, those two facts are the entire argument, and they do not resolve cleanly.
What the Disrupt debate was actually about
TechCrunch reports that founders at Disrupt 2026 spent real stage time arguing open versus closed AI. I’ve watched enough of these panels to know they tend to collapse into ideology within ten minutes. Open-source people talk about control and cost. Closed-model people talk about capability and support contracts. Everybody nods. Nobody changes their stack.
The useful version of this conversation is narrower and more boring: which failure mode can your company survive?
If you build on a closed API, your failure mode is dependency. Pricing changes, deprecations, rate limits, and model updates that quietly shift your output quality. You don’t get a vote. If you build on open weights, your failure mode is ownership. You now run inference, you tune it, you patch it, and when it degrades at 2 a.m. the on-call person is someone on your payroll.
Benchmark parity doesn’t erase that difference. It just means you can no longer use “the open models aren’t good enough” as a reason to skip the decision.
The parity claim deserves a closer look
Benchmarks closing is real and it matters, but I’d treat it as a floor, not a ceiling. Benchmarks test the things that are easy to score. Anthropic’s own breakdown is the tell here: the category that improved most dramatically was open-ended problems, moving from roughly 26 percent to 91 percent, with overall figures landing around 88 to 92 percent by September 2026. Open-ended work is exactly what resists a clean score. It’s also exactly what most agent products are selling.
So when someone tells me open and closed are now equivalent, my question is: equivalent at what? Structured tasks with verifiable answers, probably yes. Messy, multi-step, judgment-heavy work where the shape of a good answer isn’t known in advance? That’s where I’d want to test rather than assume. Run your own evals on your own data. The published numbers are marketing for both camps.
Why the “Cursor for X” idea changes the math
Foundation Capital’s read on 2026 is that “Cursor for X” becomes the default architecture for knowledge work — legal, finance, marketing, sales, operations. Tools that blend open-ended exploration with constrained precision, and that don’t feel like a chat window.
That framing is the most practically useful thing in this entire debate, because it implies you’re not making one model choice. You’re making two.
The exploration half of the product wants the strongest reasoning you can get, because that’s where users forgive latency and cost in exchange for a good answer. The precision half — extraction, classification, formatting, validation, the thousand small calls that run constantly — wants cheap, fast, predictable, and privately hosted. Those are different procurement decisions with different winners.
Most teams I’ve looked at this year are already doing this without framing it as a strategy. A frontier API for the hard path, an open model for the high-volume path. That’s not fence-sitting. It’s just matching the tool to the job, and it’s the only approach that survives either camp changing the terms on you.
What I’d actually tell a founder
- Decide which parts of your product are allowed to depend on a vendor. Write it down. That’s your real architecture document.
- Build your routing layer early, even if it points at one provider today. Swapping models should be a config change, not a rewrite.
- Run your own evals before trusting parity claims. Your users don’t live on a benchmark.
- Price the operational cost of self-hosting honestly, including the engineer who maintains it. Open weights are free; running them is not.
- If your product’s value is open-ended judgment, test the frontier models hard. That’s where the remaining gap seems to live.
At AI House Davos 2026, one speaker described European founders building on top of open AI and moving very fast. That’s the part I find most telling. The speed isn’t coming from picking the correct side of a philosophical argument. It’s coming from teams that stopped treating the question as binary and started treating it as a routing problem.
The open-versus-closed debate makes for a good panel. It makes for a bad roadmap.
🕒 Published: