DeepSeek Just Made Its 75% Off-Peak Discount Permanent. The Price War Is Real Now.

DeepSeek just turned a “try it out” promo into a permanent line item. The Chinese AI lab’s 75% off-peak discount — once a limited campaign — is now baked into its pricing forever. This isn’t a marketing stunt. It’s a shot at the entire economics of flagship LLM access.

What “permanent” actually means

The off-peak window runs 16:30 to 00:30 UTC. Hit the API during that stretch and you pay 25% of the standard rate on DeepSeek’s flagship model. Indefinitely.

To put a number on it: if you were paying around a dollar per million tokens, off-peak now costs roughly thirty cents. DeepSeek was already the cheapest serious option on the market. They just dropped the floor again — and nailed it to the ground.

Why permanence is the real news

A limited promo is bait. You can’t build infrastructure around something the vendor might yank next quarter. Permanent pricing is something a CTO can actually plan against.

That’s the unlock. Engineering teams can now redesign nightly batch jobs, data pipelines, and async workloads to deliberately land in DeepSeek’s discount window. It moves from “neat hack” to “line on the architecture diagram.” Once that decision gets made, switching back is painful — which is exactly the lock-in DeepSeek wants.

The strategic split with the West

OpenAI, Anthropic, and Google are climbing the ladder — pricier flagships, longer context, premium tiers for reasoning. DeepSeek is doing the opposite: performance is good enough, win on cost.

What makes this dangerous is which model is discounted. Most AI price cuts so far have applied to older or distilled models — the flagship stayed premium. DeepSeek just put its top-tier model on permanent sale. That breaks the unwritten rule that frontier capability commands frontier pricing.

Competitors are stuck. Match the cut and torch margins on the only SKU that funds R&D. Hold the line and watch cost-sensitive customers — startups, batch-heavy enterprises, anyone outside the Fortune 500 — quietly migrate.

Who actually benefits

The clear winners are async workloads. Overnight batch inference, large-scale data labeling, synthetic data generation, document processing — anything where latency doesn’t matter — can now run at a quarter of the cost by scheduling around the off-peak window.

The clear non-winners: real-time chatbots, coding copilots, anything that has to answer at 3pm New York time. Those workloads sit squarely in the expensive window. Expect “workload scheduling” to become a real discipline in AI infra — the same way spot instances reshaped how teams thought about EC2 a decade ago.

The takeaway

DeepSeek isn’t running a sale. They’re restructuring how the industry prices flagship intelligence, and forcing every competitor to either follow them down or defend a premium that’s getting harder to justify. The interesting question isn’t whether OpenAI or Anthropic will respond — it’s whether they can without breaking their own unit economics. If your team runs anything overnight, you have a new line item to revisit this week.