Hey 👋
Saw the tweet making the rounds about deepseek v4 being 35x cheaper than opus on input and 178x cheaper on cached tokens, and was sure it was hyperbole. Pulled the numbers anyway because i had nothing better to do.
V4-Pro input at standard rate is $0.145 per million. Claude Opus 4.7 input is around $5 per million. That is 34x. Round to 35x and the tweet is right. I genuinely thought someone had moved a decimal point.
Then there is the 75% promotional discount running through end of May, which takes V4-Pro input to $0.036 per million. Now we are at 138x cheaper than Opus.
Cache hits is where it gets funny. They cut cache hit pricing across the entire API to one-tenth of what it was. V4-Pro cache hit is now $0.0036 per million. Opus cached is around $0.625 per million. That is 173x. Once again, the tweet undersold it slightly.
The honest part nobody mentions, deepseek themselves admit V4 is three to six months behind GPT-5.4 and Gemini 3.1 Pro on capability. So you are not getting frontier quality at frontier-divided-by-178 prices. You are getting last-summer's frontier quality at a price that makes openai's billing page look like a luxury good.
For agentic loops where most input is cached system prompts and tool definitions, the cache hit discount is the actual story. The system prompt you reuse across every call is now basically free.
Anyone running deepseek v4 in prod yet, what is breaking? Curious whether the 1M context window holds up under real workloads or if it is the usual "1M but useful only to 200k" situation.
[link] [comments]