The corridor closes
Frontier API prices have fallen 430×. Open-weight alternatives fell faster. The gap between what a frontier provider charges and what an open-weight equivalent costs has collapsed from $29 to under $2—and OpenAI now spends $1.35 for every $1.00 it earns.
Open-weight models didn't just erode frontier pricing—they restructured what frontier pricing can be. In 2023, GPT-4 commanded $30 per million tokens while the best open-weight alternative cost $0.70. The corridor was $29.30 wide. By 2026, GPT-5 nano reaches $0.60, but DeepSeek V3.2 serves equivalent performance at $0.07. The corridor is now $0.53. And OpenAI’s own financials show the cost floor sits above what the market will pay: the company spends $1.35 for every dollar it earns. The corridor hasn’t just closed—at current prices, frontier providers are operating below their own cost floor.
Sources & methodology
Frontier API prices: OpenAI published pricing at each model launch. GPT-3 Davinci (Nov 2021) $60/M combined; GPT-4 (May 2023) $30/M input; GPT-4o (May 2024) $2.50/M input; GPT-5.2 (2025) $1.75/M input; GPT-5 nano (2026) $0.60/M input. Open-weight hosted pricing: Llama 2 (2023) ~$0.70/M via major hosting providers; Llama 3.1 (2024) ~$0.50/M; DeepSeek V3 (2025) $0.14/M; DeepSeek V3.2 (2026) ~$0.07/M (10–30× below GPT-5-level frontier pricing, midpoint). OpenAI unit economics: company reported $1.35 cost per $1.00 revenue (2026); projected $14B loss on ~$10B revenue. Cost floor estimated as current price × 1.35. Sources: OpenAI pricing; TokenMix pricing history; AI Automation Global (OpenAI unit economics).