Closed AI models now dominate the price-performance frontier across nearly all price points, according to analysis from Village’s LLM Atlas, which maps the strongest AI model available at each price level.
Open-weight models currently hold the best-value position at only 3% of budgets, down dramatically from 41% in June. This represents the lowest share measured since the end of 2024. The few remaining open-source winners at competitive prices include Google’s Gemma 4 E4B and Xiaomi’s MiMo-V2.6-Flash, which dominate only the cheapest tiers.
The shift accelerated following recent releases from major closed labs. OpenAI’s GPT-6 Luna (September) and GPT-6.1 Terra (October) reclaimed most budget categories that open models previously dominated, alongside Anthropic’s Claude Haiku 5.5. OpenAI’s entire GPT-6 and GPT-6.1 family holds 11 of the frontier’s 21 price-performance points, according to the analysis.
Historically, open models have captured value through sporadic but dramatic surges. DeepSeek-V3.2-Exp was the best buy across nearly half of all budgets when it launched in September 2025, and DeepSeek V4 Flash previously held nearly a third of budgets in June. However, each peak has been followed by declines of about half or more as closed labs shipped cheaper alternatives within weeks.
Village identified three structural reasons closed labs maintain price-performance advantages. First, they can distill their own frontier models in-house—training cheaper versions on outputs from stronger proprietary models—while open labs must train from weaker existing models due to licensing restrictions. Second, closed labs release entire model families covering many price points simultaneously, ensuring comprehensive market coverage. Third, companies like OpenAI, Anthropic, and Google operate inference at enormous scale, with Google running on its own chips, potentially lowering per-task costs compared to open-model labs serving smaller traffic volumes.
The analysis uses Village’s LLM Atlas, which estimates capability scores from public benchmarks and calculates costs per task from evaluators’ API runs on a log scale across 201 budgets. Costs are relative to GPT-5 at high effort. The methodology acknowledges that open models can be cheaper when run on personal hardware rather than through hosted APIs, and that tiny open models still win at the absolute cheapest end when cost ranges are widened further.
Key facts
- Open-weight models now hold the best-value position at only 3% of budgets, down from 41% in June
- OpenAI’s GPT-6 and GPT-6.1 families hold 11 of the frontier’s 21 price-performance points
- Closed labs can train cheaper model tiers from frontier models in-house, while open labs face licensing restrictions
- DeepSeek-V3.2-Exp held nearly 50% of budgets the week it launched in September 2025 before its share declined