Open models hold just 3% of price-performance budgets in Village's LLM Atlas

Closed models hold nearly every budget

Village's LLM Atlas maps the price-performance frontier for language models: at every price point, it identifies the most capable model available. According to the latest reading published on Village's blog and surfaced on Hacker News, closed models now dominate that frontier. An open-weight model is the best-value pick at only 3% of budgets — the lowest share Village has measured — and nearly all of those wins sit at the very cheap end of the range, held by Google's Gemma 4 E4B and Xiaomi's MiMo-V2.6-Flash.

The collapse has been fast. Village reports that open models held 41% of budgets in June, with DeepSeek V4 Flash alone the top pick across nearly a third of budgets and Zhipu's GLM-5.2 just above it. Two OpenAI releases — GPT-6 Luna in September and GPT-6.1 Terra in October — reclaimed most of that territory, alongside Anthropic's Claude Haiku 5.5.

A recurring pattern of bursts

Village notes that open models have led on value before, but only in bursts. At the end of every quarter since 2024, open models held at least 6% of budgets. The peaks came from single releases: DeepSeek-V3.2-Exp was the strongest option at nearly half of all budgets during the week it launched in September 2025. None of those peaks lasted — after each one, the open share fell by roughly half or more by the following quarter as closed labs shipped cheaper tiers.

The current dynamic, as Village describes it, is that a strong, cheap open release takes a wide band of budgets until the next closed model undercuts it. The newest closed releases have been both cheaper and stronger than anything open, so the open band has shrunk to the very bottom of the price range.

Why closed labs may be winning

Village offers three explanations that fit the pattern:

1. Distillation from frontier models. A lab with the strongest model can train cheaper ones on its outputs in-house, and cheap closed tiers have followed their flagships within weeks — Claude Haiku 5.5 arrived two weeks after Claude Opus 5.5. Open labs must instead start from their own weaker models, since closed labs' terms generally bar training competitors on their outputs.
2. Whole model families. OpenAI's GPT-6 and GPT-6.1 lines, spanning Luna and Terra up to Sol and Astra, hold 11 of the frontier's 21 points between them, so a single generation covers nearly every price.
3. Cheaper serving. OpenAI, Anthropic and Google run inference at enormous scale — Google on its own chips — which can lower the cost of each task. Most open-model labs serve far less traffic.

Caveats in the methodology

The Atlas estimates a capability score for each model and reasoning setting from public benchmark results, and a cost per task from evaluators' runs, with costs measured relative to GPT-5 at high effort. For each date, it selects the most capable model at each of 201 budgets spread evenly on a log scale from 1% to three times GPT-5's cost. The scores behind the study date from October 31, 2026.

Costs reflect what evaluators paid through hosted APIs, and Village acknowledges that open models can be cheaper to run on your own hardware. The pattern holds under two robustness checks: restricting the analysis to models scored on at least 10 benchmark families, or narrowing the budget range to 3%–1x GPT-5's cost. Widening the range to 0.3%–10x raises the open share to 16%, because tiny open models still win at the very cheapest end.

Why it matters

Anyone choosing a model by price and capability — engineers picking an API, teams budgeting production workloads — faces a frontier where closed models are the rational buy at almost every spend level. The data suggests that through hosted APIs, open weights no longer guarantee savings at equivalent quality, and that open labs' advantage now appears only in short bursts after major releases, at the extreme low end of cost, or when self-hosting changes the economics entirely. It also points to a structural constraint: with closed labs' terms blocking training on frontier outputs, open labs may keep losing ground between their own releases.