Glonce

Village's LLM Atlas: open-weight models win only 3% of budgets, down from 41%

researchThe Glonce editorsHacker News🔥66Glonce rating66Impact48Novelty45Relevance95Surprise45Players45Credibility60
Village's LLM Atlas: open-weight models win only 3% of budgets, down from 41%

Village publishes the LLM Atlas, which maps the price-performance frontier: the strongest AI model at every price point. In a new study using scores from October 31, 2026, it reports that closed models now hold almost every budget. An open-weight model is the best buy at only 3% of budgets, and nearly all of those are at the cheap end, held by Google's Gemma 4 E4B and Xiaomi's MiMo-V2.6-Flash.

In June the figure was 41%. DeepSeek V4 Flash alone was the best buy across nearly a third of budgets, with Zhipu's GLM-5.2 just above it. Two OpenAI releases took most of that back: GPT-6 Luna in September and GPT-6.1 Terra in October now cover the budgets open models used to win, alongside Anthropic's Claude Haiku 5.5.

The study says open models have led on value before. Today's share is the lowest Village has measured; at the end of every quarter since 2024, open models held at least 6% of budgets. The earlier 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 the 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 authors summarise the pattern this way: open models win on value in bursts, when a strong, cheap release (usually from DeepSeek) takes a wide band of budgets until the next closed model undercuts it. The latest closed releases have been both cheaper and stronger than anything open, so the band has shrunk to the very bottom of the price range.

The study offers three explanations that it says fit the pattern. First, closed labs can distill their own frontier models: a lab with the strongest model can train cheaper ones on its outputs in-house, and today's cheap closed tiers followed their flagships within weeks, as Claude Haiku 5.5 did two weeks after Claude Opus 5.5. Open labs have to start from their own, weaker models, since closed labs' terms generally bar training competitors on their outputs. Second, closed labs release whole families: OpenAI's GPT-6 and GPT-6.1 models, from Luna and Terra up to Sol and Astra, hold 11 of the frontier's 21 points between them, so one generation covers nearly every price. Third, closed labs may serve more cheaply: OpenAI, Anthropic and Google run inference at enormous scale, Google on its own chips, which can lower what each task costs, while most open-model labs serve far less traffic.

On method, the Atlas estimates a capability score for each model and reasoning setting from public benchmark results. A separate cost model estimates cost per task from evaluators' runs. Costs are relative to GPT-5 at high effort. For each date, the study takes every model released by then and finds the most capable one at each of 201 budgets, spread evenly on a log scale from 1% to three times GPT-5's cost. Costs reflect what evaluators paid through hosted APIs, and open models can be cheaper to run on your own hardware. Restricting to models scored on at least 10 benchmark families, or narrowing the budgets to 3% to one times GPT-5's cost, gives the same pattern. Widening them to 0.3% to ten times GPT-5's cost raises today's open share to 16%, because tiny open models still win at the very cheapest end.

Key facts

  • An open-weight model is the best buy at 3% of budgets as of October 31, 2026, down from 41% in June; nearly all of the 3% sits at the cheap end, held by Gemma 4 E4B and MiMo-V2.6-Flash.
  • In June, DeepSeek V4 Flash alone was the best buy across nearly a third of budgets; OpenAI's GPT-6 Luna (September) and GPT-6.1 Terra (October) now cover those budgets, alongside Claude Haiku 5.5.
  • Today's open share is the lowest Village has measured; open models held at least 6% at the end of every quarter since 2024, and after each peak the share fell by about half or more by the next quarter.
  • OpenAI's GPT-6 and GPT-6.1 models hold 11 of the frontier's 21 points; the authors suggest distillation, whole-family releases and cheaper serving at scale as possible reasons for the closed labs' lead.
  • Costs are relative to GPT-5 at high effort and based on hosted APIs; widening the budget range to 0.3% to ten times GPT-5's cost lifts today's open share to 16%.

Why it matters

Open-weight models have often been the cheap way to get strong capability. This study says that is no longer true across most of the price range. The open share of budgets fell from 41% in June to 3% as OpenAI's GPT-6 Luna and GPT-6.1 Terra arrived, and it is the lowest figure Village has measured. The history matters too: DeepSeek-V3.2-Exp led at nearly half of all budgets in September 2025, yet the open share fell by about half or more within a quarter of each peak. The authors read this as a cycle in which open releases win in bursts until a closed model undercuts them.

Who it affects

Teams choosing a model for a given price point are the direct audience, since the Atlas maps the strongest model at each budget. Open-model labs such as DeepSeek, Zhipu, Google (Gemma) and Xiaomi are affected because their position now sits at the very bottom of the price range. OpenAI, Anthropic and Google are the closed labs the study credits with the latest cheaper and stronger releases.

How to use it

Read the result as a guide to hosted-API value, not as a verdict for every deployment. The costs reflect what evaluators paid through hosted APIs, and the source says open models can be cheaper to run on your own hardware. At the very cheapest end, tiny open models such as Gemma 4 E4B and MiMo-V2.6-Flash still win. The Atlas updates as new results arrive, so the picture can shift when the next open release lands.

How solid is it

The analysis comes from Village's own LLM Atlas and is not peer-reviewed; the source names no individual authors. Capability scores are estimated from public benchmark results, and costs come from a separate cost model based on evaluators' runs. The result held when Village restricted the data to models scored on at least 10 benchmark families, or narrowed budgets to 3% to one times GPT-5's cost. The three explanations for the closed labs' lead are offered as possibilities, and the source gives no evidence that cheaper closed tiers were actually distilled from flagships.

Risks and caveats

Costs are relative to GPT-5 at high effort, not in dollars, and no dollar prices are given. Self-hosted cost of open models is not measured. The headline share depends on the budget range chosen: 201 budgets from 1% to three times GPT-5's cost give 3%, while widening to 0.3% to ten times raises it to 16%. The explanations (distillation, family releases, serving scale) are hypotheses that the authors say fit the pattern, not demonstrated causes.

“Open models win on value in bursts: a strong, cheap release, usually from DeepSeek, takes a wide band of budgets until the next closed model undercuts it.”

— Village, Closed AI models now dominate the price-performance frontier