DeepSeek shifts pricing: double peak rates, cheap cache

According to a post by X user Teortaxes, DeepSeek's pricing model changed on September 10, when the "Intermediate" V4.1 version expired and, in their words, would be "succeeded by who knows what." The change rolls back read and cache-miss prices to their pre-hike levels, while output token pricing stays at a premium.

New rate structure: off-peak, a cache hit costs 0.02 yuan, a cache miss 1 yuan, and output 4 yuan per million tokens. Peak hours double those rates — 0.04 yuan per hit, 2 yuan per miss, 8 yuan per output. Effective September 10.

The interesting part isn't the numbers themselves but the structure. DeepSeek puts read costs on the floor — 0.02 yuan per cache hit — and shifts revenue to output. Cache misses return to 1 yuan, the level before the temporary increase, suggesting the earlier experiment in raising read prices didn't stick. Output at 4 yuan per million tokens off-peak remains relatively expensive compared with comparable Chinese rivals.

Unofficial source, official uncertainty: the information comes from a tweet by a user who identifies as a longtime fan ("DeepSeek 铁粉 2023 - ∞"), not from an official announcement. Teortaxes explicitly notes the next version is unknown ("succeeded by who knows what"), making the current pricing temporary by definition. DeepSeek itself had not published updated documentation on its developer site as of the report.

What it means for developers: for teams building on aggressive caching, the near-zero hit price and 1-yuan miss are manageable. The problem is output — 4 yuan per million tokens off-peak, 8 at peak — adds up fast in long chat sessions or tool-calling chains. Without official specs on cache limits, TTL, or the definition of "peak hours," real-world cost is hard to calculate. For now, developers are running on pricing no one has signed off on.