The era of open-weight models dominating the value proposition for general-purpose AI inference appears to be over, at least for now. New data from Village’s LLM Atlas, a benchmarking framework that maps the strongest AI model at every price point, reveals that closed models now hold nearly every budget tier. As of late October 2026, open-weight models are the best buy in only 3% of measured budgets, a drastic collapse from the 41% share they commanded in June of the same year. This shift marks the lowest value share for open models since Village began tracking the frontier in 2024.

The Collapse of the Open Value Proposition

The decline is driven by aggressive releases from closed labs that have simultaneously increased capability and lowered costs. In June, DeepSeek V4 Flash and Zhipu’s GLM-5.2 were the go-to choices for cost-sensitive developers, with DeepSeek alone covering nearly a third of all budgets. That advantage evaporated quickly. OpenAI’s GPT-6 Luna, released in September, and its successor GPT-6.1 Terra, launched in October, reclaimed those budget slots. These new closed models, along with Anthropic’s Claude Haiku 5.5, offer better performance-per-dollar than any open alternative in the mid-to-high price ranges. Today, the only remaining open winners are Google’s Gemma 4 E4B and Xiaomi’s MiMo-V2.6-Flash, which hold the absolute cheapest end of the spectrum.

Why Closed Labs Are Winning on Economics

Village’s analysis suggests three structural advantages for closed labs that explain this trend. First, closed labs can distill their own frontier models. Because they own the training data and outputs, they can train cheaper, smaller models on their own flagship performance without legal restrictions. For instance, Claude Haiku 5.5 was released just two weeks after Claude Opus 5.5, effectively offering a distilled, cheaper version of the top-tier model. Open labs, conversely, are often barred from training on closed model outputs, forcing them to start from weaker baselines. Second, closed labs release entire families of models. OpenAI’s GPT-6 and GPT-6.1 series hold 11 of the 21 points on the current price-performance frontier, allowing a single generation to cover nearly every price point from budget to premium.

Infrastructure Scale and Cost Efficiency

The third factor is raw infrastructure efficiency. OpenAI, Anthropic, and Google operate inference at enormous scale, with Google leveraging its own custom TPU chips to lower the cost per task. Most open-model labs serve significantly less traffic and lack this vertical integration, making it difficult to compete on price even when algorithmic efficiency is high. Village notes that while open models can sometimes be cheaper to run on private hardware, the hosted API costs—which most developers pay—have decisively shifted in favor of closed providers. The data shows that when restricting the analysis to models scored on at least 10 benchmark families, the pattern of closed dominance holds firm, with open models only gaining a 16% share if the budget range is widened to include the tiniest, least capable models at the extreme low end.

Key Takeaways

  • Open-weight models now hold only 3% of the price-performance frontier, down from 41% in June 2026.
  • OpenAI’s GPT-6 Luna and GPT-6.1 Terra, along with Anthropic’s Claude Haiku 5.5, have displaced DeepSeek and Zhipu models from mid-range budgets.
  • Closed labs benefit from distilling their own frontier models and releasing full product families that cover all price tiers.
  • Infrastructure scale and custom silicon give closed labs a cost-per-task advantage that open labs struggle to match via hosted APIs.
  • The only remaining open-model strongholds are the absolute cheapest tiers, held by Gemma 4 E4B and MiMo-V2.6-Flash.

The Bottom Line

The open-weight movement has lost its economic argument for mainstream developers. Unless open labs can solve the infrastructure scaling problem or find a way to legally leverage closed-model outputs for training, they are being pushed into a niche of ultra-low-cost experimentation rather than serious production workloads.