Open-Weight AI Models Seize Token Lead, but Proprietary Still Make the Money

Open-weight (downloadable) models have surged to process 56% of production tokens routed through Vercel’s AI Gateway in August, up from 36% in July and under 10% in December 2025, with a single day hitting 62% open-weight traffic. Despite dominating volume, open models accounted for only 14% of estimated spending because open-weight inference costs are far lower: closed-weight tokens cost roughly 7.8× more on average (open tokens about 87% cheaper). Capability differences have narrowed - Mozilla’s State of Open-Source AI 2026 finds a roughly 3.3-point gap, with open models near parity on coding, instruction-following and general knowledge - prompting enterprises to shift heavy production workloads to open models and reserve proprietary systems for value-justified tasks.

Price competition is driving adoption and pushing gateway token prices down (average price per token fell 23.2% in August; high-usage teams saw median per-token price drop 7.6%). Proprietary vendors still capture the bulk of spending - Anthropic models represented about 64% of estimated August spend - yet customers trade down within vendor lineups when cheaper options suffice: Anthropic’s Fable 5 spending share fell from 13.2% to 4.9% while Opus 5 rose to 22.5% (about half Fable’s per-token cost). Vercel’s figures cover anonymized gateway traffic and use list-price estimates, so they exclude direct API, self-hosted or enterprise-contract nuances, but the clear takeaway is that pricing, not frontier capability, is increasingly the decisive factor.