Closed AI models dominate the price-performance frontier

Closed-weight AI models now offer the best capability for almost every tested budget. In June 2026, open-weight models were the strongest option at 41% of price points; by October 31, they led at just 3%, the lowest share measured since 2024. Their remaining wins are concentrated at the cheapest end, chiefly Google’s Gemma 4 E4B and Xiaomi’s MiMo-V2.6-Flash. Open models have repeatedly surged after standout releases - DeepSeek-V3.2-Exp briefly led across nearly half of budgets in September 2025 - but their advantage then shrank as closed labs introduced cheaper alternatives. GPT-6 Luna and GPT-6.1 Terra, alongside Anthropic’s Claude Haiku 5.5, have reclaimed much of the territory open models held earlier in 2026.

Three factors may explain the shift: leading closed labs can train inexpensive models from their own stronger systems, release full model families spanning many price levels, and serve requests at lower cost through greater infrastructure scale. The comparison estimates capability from public benchmarks and task costs from evaluator runs, relative to GPT-5 at high effort. It assesses the best model at 201 budgets ranging from 1% to three times that reference cost. The results are based on hosted API prices, so open models may be cheaper when self-hosted. The overall pattern also persists under narrower budget ranges and stricter benchmark coverage, though extending the range to extremely low budgets raises open models’ share to 16%, reflecting their continued strength among tiny, inexpensive systems.