Open-weight models win just 3% of price points, down from 41% in June

An open-weight model is the best buy at 3% of budgets, down from 41% in June. That is the headline number in [Village’s LLM Atlas update](https://village.dev/blog/closed-ai-models-now-dominate-the-price-performance-frontier/), published October 30, and it is the sharpest collapse in open-model value share the tracker has recorded. The Atlas maps the price-performance frontier: at each of 201 budgets spread on a log scale from 1% to three times GPT-5’s cost, it finds the most capable model available as of a given date. In June, DeepSeek V4 Flash alone was the best buy across nearly a third of those budgets, with Zhipu’s GLM-5.2 just above it. Today almost every point belongs to a closed model.

The reversal is not a slow drift. Village attributes most of it to two OpenAI releases: GPT-6 Luna in September and GPT-6.1 Terra in October, which now cover the budgets open models used to win, alongside Anthropic’s Claude Haiku 5.5. The open-weight share that remains is concentrated at the very cheap end, held by Google’s Gemma 4 E4B and Xiaomi’s MiMo-V2.6-Flash. Widen the budget band to 0.3% through ten times GPT-5’s cost and today’s open share rises to 16%, because tiny open models still win at the floor. The pattern holds when Village restricts to models scored on at least 10 benchmark families, or narrows the budgets to 3% through one times GPT-5’s cost.

The frontier is a moving target, and open labs keep missing it

Open models have led on value before. Village notes that at the end of every quarter since 2024, open models held at least 6% of budgets, and the peaks came from single releases. DeepSeek-V3.2-Exp was the best buy at nearly half of all budgets the week it launched in September 2025. None of those peaks lasted. After each one, the open share fell by about half or more by the next quarter as closed labs shipped cheaper tiers.

The honest read is that the open-weight camp has been running a release-by-release insurgency against a supply chain it does not control. DeepSeek can drop a model that undercuts everything on the market for a few weeks. Then OpenAI ships a Luna or a Terra, Anthropic ships a Haiku, and the band collapses back to the bottom of the price range. The Atlas data makes the asymmetry visible: the entire open-weight position now rests on two models at the cheapest budgets, not on a broad competitive front.

Why the closed labs can win on price at all

Village offers three explanations, and the first is the most consequential for how the industry is structured.

Closed labs can distill their own frontier models. A lab holding the strongest model can train cheaper ones on its outputs in-house, and the cheap closed tiers have followed their flagships within weeks. Claude Haiku 5.5 arrived two weeks after Claude Opus 5.5. Open labs have to start from their own, weaker models, because closed labs’ terms generally bar training competitors on their outputs. That is a structural advantage written into the terms of service, not a temporary lead. If the strongest model is closed, the strongest distillation target is closed too, and the open lab is left distilling its own second-best work.

The second explanation is portfolio breadth. OpenAI’s GPT-6 and GPT-6.1 families, from Luna and Terra up to Sol and Astra, hold 11 of the frontier’s 21 points between them. One generation covers nearly every price. The third is inference cost. OpenAI, Anthropic and Google run inference at enormous scale, Google on its own chips, which can lower what each task costs. Most open-model labs serve far less traffic and pay more per token for the privilege.

The entire open-weight position now rests on two models at the cheapest budgets, not on a broad competitive front.

The caveat that matters, and the one that does not

Village is careful about its own methodology, and the caveats deserve repeating. Costs reflect what evaluators paid through hosted APIs, and open models can be cheaper to run on your own hardware. That is a real gap. A team with spare GPUs and a willingness to operate its own inference stack can beat the hosted price on an open model. But the Atlas measures the price a buyer actually pays for a hosted endpoint, and that is the number most teams face when they pick a model for a product. Self-hosting is a capital expense and an operational burden, not a discount available to everyone.

The scores are from October 31, 2026, and Atlas updates as new results arrive. The study estimates capability from public benchmark results and cost per task from evaluators’ runs, with costs relative to GPT-5 at high effort. That is a defensible construction, and the fact that the pattern survives the robustness checks Village ran is the part that should worry open-model advocates.

What this means for builders

If you are building an AI product and your selection process still assumes open weights are the value play, the Atlas says that assumption is now wrong at almost every price point. The best buy at your budget is probably a closed model, and the gap is widest in the middle of the market where most production workloads live. The practical move is to re-run your cost-per-task comparison against the current frontier rather than the one you benchmarked in June.

The deeper question is whether the open-weight camp can break the distillation asymmetry. Village’s data suggests the current model, a strong cheap release every few months that gets undercut within a quarter, produces shrinking returns. DeepSeek V4 Flash and GLM-5.2 were real wins. They did not hold. Unless an open lab ships a frontier-class model that closed labs cannot distill around, or unless self-hosting costs fall far enough to change the hosted-API comparison, the 3% figure is a floor that keeps getting lower. Village will publish the next Atlas update as new benchmark results land, and the number to watch is whether the open share recovers above the 6% quarterly floor it held for two years.