Closed AI models now occupy nearly all of the price-performance frontier measured by Village's LLM Atlas, according to a new analysis from the developer-platform company. Village reports that an open-weight model was the strongest option at only 3% of the budget levels it tested, down from 41% in June.
The finding comes from a modeled comparison rather than a universal measure of model value. Atlas combines public benchmark results into a capability estimate and separately calculates an expected cost per task from evaluator runs. Village then identifies the highest-scoring model available at each of 201 budgets distributed on a logarithmic scale, from 1% to three times the cost of GPT-5 at high reasoning effort. The scores used in the analysis were current as of October 31, 2026.
Village says the pattern has historically shifted in bursts. A capable and inexpensive open model, often from DeepSeek, can take a broad range of budget points when released. Subsequent proprietary releases can reclaim those positions if they pair higher scores with lower hosted prices. In the latest snapshot, Village says that open models retained an advantage mainly at the cheapest end of the tested range.
The result depends materially on the selected price window. When Village expanded the range from 0.3% to ten times GPT-5's cost, the current open-model share rose to 16%, largely because very small open models performed well at the lowest budgets. The company says the broad pattern remained when it limited the analysis to models covered by at least 10 benchmark families or narrowed the main budget range.
The cost calculation reflects hosted API prices paid by evaluators. It does not measure every way an organization might deploy an open model. Companies using their own hardware, committed capacity or highly optimized inference software may reach different conclusions. Likewise, benchmark aggregation cannot capture every workload-specific requirement, privacy constraint or benefit of controlling model weights.
Atlas also points to rapid price compression. Village estimates that a task costing 79 cents with GPT-5 at its launch can now be completed for roughly 2 cents with a newer model. The analysis says newer releases are being matched at lower prices more quickly than older ones were, suggesting that the economic frontier is changing faster even as closed providers currently lead it.
Village's geographic comparison reports a separate trend in open weights: China's leading open model has scored above the leading US open model continuously since September 2024, under its methodology. That claim, like the price frontier itself, is best treated as a snapshot of Atlas's selected benchmarks and cost assumptions rather than a definitive ranking for every production use case.
