Open weights sit at 58 percent of the frontier and 44 percent of the price
Two long-horizon leaderboards published this month agree on how far back the open models are, and disagree completely on what that is worth.
Two people who follow open-weights models for a living said opposite things about them within ten days, and both were right.
Ethan Mollick wrote on 9 September 2026 that open models are further from the frontier than we have seen in some time. Yuchen Jin, who runs open models professionally, had written days earlier that there is no default winner any more. They are measuring different columns, and the arithmetic that separates them has been sitting on two public leaderboards all week.
Start with score.
On Vercel's DeepsecBench, which scores models on finding real vulnerabilities in application code, the best open-weights configuration reaches 58 percent of the leader's score. Every other open model on the board sits further back.
Andon Labs' Vending-Bench 2, which drops an agent into a simulated vending business for a year and scores it on the closing bank balance, lands in the same place from a completely different direction: the best open model finishes at 53 percent of the leader. Two evaluations with nothing in common — different harnesses, different tasks, different scoring units — agree to within five points on how far back open weights are.
So Mollick is probably right about capability. Now price it.
The same run, in money
That DeepsecBench run cost the frontier model $63.70 and the open one $16.22. Per point of score, the open model is about 2.3 times cheaper, and that ratio is roughly why you keep finding open weights in the parts of a stack nobody puts on a slide. On Ollama's cloud the small GLM variant runs at $0.15 per million input tokens against $10 for Astra — different tiers of model, yes, and the ratio is still sixty-seven to one.
There is also one board where the open model simply won. In Vending-Bench Arena round twelve, three agents competing at the same simulated location, GLM-5.3 finished second on $7.8k and Claude Fable 5.1 came last on $5.7k. An open-weights model beat a frontier lab's newest release at running a business for a year.
What about running one on your own hardware? Tom's Hardware spent a day on that with Qwen 3.8 27B, roughly 17GB of four-bit weights, and found that memory capacity alone cannot overcome software and inference-engine bottlenecks. The single-card frontier is still not here — not for lack of weights, but for lack of an engine that can feed them.
The gap is widening in score and narrowing in money
Most of the argument people are having is two groups reading different columns of the same table. Andon publishes a regression on exactly this: the Chinese frontier gaining $1,047 a month on Vending-Bench against the Western frontier's $822, a lag of about 111 days, with crossover projected for October 2027.
We would take the other side of that date. The fit rests on one data point that broke the Western trend upward, and a linear projection drawn through a step change is, as far as we can tell, the most reliable way to be wrong in public. Sometime before then an open-weights model will crack the top three of DeepsecBench. If it does before Christmas, we were too cautious here, and we will say so.
Until that happens, what you are actually choosing is 58 percent of the answer at 44 percent of the cost. For most work, most of the time, that is not a hard call — which is the part Mollick's post, correct as it is about capability, does not price.
