DeepSeek

DeepSeek sets the floor other price sheets are read against. On 31 July 2026 it published deepseek-ai/DeepSeek-V4-Flash-0731 on Hugging Face under the MIT License — weights, commercial use, no threshold clause. The OpenRouter row for it lists at $0.06per million tokensopenrouter-models, last checked 2026-09-14 input with an Artificial Analysis intelligence index of 34.5openrouter-models, last checked 2026-09-14.

Put that next to what the same catalog charges on the other side of the price sheet. OpenAI's openai/gpt-5.4, its flagship in March 2026, lists at $2.50per million tokensopenrouter-models, last checked 2026-09-14; Google's google/gemini-3.5-flash, from May, at $1.50per million tokensopenrouter-models, last checked 2026-09-14. Same snapshot, and both are more than an order of magnitude above the DeepSeek row. Simon Willison, who runs models against his own tasks rather than against leaderboards, wrote that the release's pricing means this "may currently be the best value-per-intelligence model out there".

There is a stated mechanism behind the number, not just a price decision. The Flash model is a sparse mixture of experts: OpenRouter's own listing describes roughly 13 billion parameters active out of 284 billion total, so what a request costs to serve tracks the active path rather than the published size. DeepSeek's V4 line arrived as a preview on 24 April 2026, went final for Flash on 31 July, and for Pro on 13 August — three months from preview to two shipped sizes, with the weights of both published rather than described.