Is OpenAI's new pricing really beating DeepSeek, or is there a catch?

The headline number says OpenAI just got cheaper than China's favorite budget model. Look one line deeper and it's not that simple.

Here's the number that's making the rounds. OpenAI's new cheap model, Luna, charges ten cents per million input tokens. DeepSeek's budget model, the one everyone points to when they want to prove open weight models beat the big labs on price, charges fifteen cents off peak. So on paper, OpenAI just underpriced China's price leader at its own game. That's new. DeepSeek has basically owned the "cheapest serious model" title for two years running.

So is DeepSeek suddenly the expensive option?

Hold on though, if DeepSeek has been the price leader this whole time, how did OpenAI just walk in and undercut it in one release?

Here's the part that doesn't show up in the headline number. Both models offer something called a cache discount, basically, if you send the same or overlapping content repeatedly, like a system prompt or a codebase an agent keeps rereading, the model doesn't have to reprocess it from scratch, and you pay less for it. DeepSeek's cache price is almost nothing, fractions of a cent per million tokens. Luna's cache price is still a real chunk of its full rate.

So if your workload is the kind that keeps hitting the same context over and over, a coding agent reading the same repo a hundred times a day, a chatbot reusing a long system prompt, DeepSeek can end up cheaper overall even though its sticker price looks higher. The fresh token race, OpenAI wins. The repeated token race, the one that actually matters for a lot of production workloads, DeepSeek still has the edge.

The bigger game being played here

Zoom out and this stops being about ten cents versus fifteen cents. The whole pitch for open weight models like DeepSeek has always been cost, you can't self host GPT-6, but you can run DeepSeek's weights yourself and skip the API entirely. If a closed, hosted model from OpenAI is now roughly as cheap and roughly as capable, that argument gets a lot weaker for anyone who was choosing open weight purely on price rather than on actually wanting to self host.

That's the real move here. Not "we built a smarter model." It's "we're making the reason you'd choose the cheaper alternative disappear."