Last week, DeepSeek quietly made something permanent that most enterprise software buyers haven’t noticed yet. A 75% discount on their best model — originally a promotional event set to expire May 31 — is now just the price. No expiration. No asterisk.
That means running 100 million output tokens through DeepSeek’s V4-Pro costs about $348 a month. The same workload through GPT-5.5 or Claude Opus costs $2,500 to $3,000. Same task. Same volume. Eight to nine times the price difference.
Most finance teams I talk to have no idea this gap exists — let alone that it just became permanent.
How We Got Here
When GPT-4 launched at the end of 2022, frontier AI cost around $60 per million output tokens. There was one tier: expensive. Then in January 2025, DeepSeek dropped R1 — a Chinese-built model that matched US frontier quality at a fraction of the cost. NVIDIA’s market cap dropped $600 billion in a single day. The market got the message: this pricing structure wasn’t a law of nature.
Over 2025, every major lab cut their mini and flash tiers — that’s why every model release now comes in small, medium, and large versions. But the headline frontier models stayed expensive, stuck around $25–$30 per million output tokens. Until now.
Why DeepSeek Can Do This
A few reasons. Their architecture is more compute-efficient — less processing per token. Their hardware runs on Huawei chips outside US export controls. And they’re backed by High-Flyer, a Chinese quantitative hedge fund that doesn’t need per-token revenue to survive the way a VC-backed US lab does. They’re not running a promotion. This is their business model.
Worth noting: DeepSeek’s models are not as good as Claude or GPT-5.5. They’re very good compared to where the industry was two years ago — but they’re not equivalent. That distinction matters a lot for how you think about using them.
The Cost Shock Nobody Modeled
Here’s a pattern I’ve seen play out a lot lately. A firm moves from personal ChatGPT subscriptions — $20 a month per person — to an enterprise token-based plan. They discover they’re now spending closer to $20 a day. Same capability on paper. Completely different economics at scale.
People who have tested this with $200 pro plans report getting the equivalent of about $2,000 worth of tokens — roughly a 10x difference between seat-based and usage-based pricing. Most organizations never modeled what scale would look like before signing. Now they’re living with it.
The pricing gap that DeepSeek just made permanent means every one of those conversations now has a benchmark to push against.
How to Actually Think About Model Selection
The useful framework here is reasoning complexity versus volume. Two axes, four quadrants.
High reasoning complexity at low volume — ambiguous inputs, competing constraints, professional liability on the output — that’s where frontier models justify their price. The additional reasoning quality is worth the premium because the stakes are high and the volume is low enough that the cost delta doesn’t compound.
High volume at low reasoning complexity is the opposite. Structured inputs, predictable outputs, lots of repetition — this is where the 7–9x price gap creates real budget impact. I ran into this recently: had a bunch of people submit budget requests with reasons for needing AI token access. Instead of running Sonnet over all of them, I used Haiku — a fraction of the cost — and got a solid first-pass summary of every submission. Right tool for the job.
The middle quadrants — high complexity at high volume — are where you need to do actual benchmarking before committing. Try both models. See if you can feel the quality difference. If you can’t, you have your answer.
The Data Residency Problem
Before you route client data to DeepSeek’s hosted API, you need to understand one thing: all data processed through their hosted API is stored on servers in China, subject to Chinese Cybersecurity Law and National Intelligence Law. Chinese authorities can compel access without user consent. This isn’t a hypothetical — it’s the documented legal structure.
The workaround is self-hosting. DeepSeek releases their model weights under an MIT license, which means you can download and deploy them yourself. AWS Bedrock and Azure AI both offer hosted versions of the open weights in US-jurisdiction infrastructure — that’s the middle path for firms that want the cost advantage without the data risk. It requires more infrastructure investment, but the cost gap still doesn’t close even accounting for that.
For internal tasks with no client data — internal research, summarizing public information, processing non-confidential documents — the hosted API may be fine. But check your firm’s data policy first. Don’t assume.
Key Takeaways
- The 7–9x pricing gap between DeepSeek and US frontier models is real and verifiable from public rate cards today.
- Match the model tier to the task — not to habit. High-volume, structured work doesn’t need frontier pricing.
- Treat AI API spend like any critical vendor contract: benchmark annually, monitor usage by task type, model what scale looks like before signing.
- DeepSeek’s hosted API sends data to China. For client-facing work, the path to those prices runs through self-hosting or US-jurisdiction infrastructure like AWS Bedrock or Azure AI.
- The Big Four AI deals with Anthropic and OpenAI are switching cost strategies as much as technology decisions — embedded workflows make per-token economics less visible over time.
Want the CPE credit? Take the full lesson on EverydayCPE and earn 0.2 CPE credits: [lesson link]
Leave a Reply