DeepSeek's models have become a go-to for developers who want near-frontier reasoning without the usual bill shock. But the official pricing page only tells part of the story: the same model can cost noticeably different amounts depending on which provider you route through, and the gap grows once you factor in caching, volume discounts, and markup.
In this post I'll break down what DeepSeek API pricing actually looks like today, how third-party gateways compare, and a few practical tricks to cut your per-million-token spend.
How DeepSeek API pricing is quoted
DeepSeek quotes prices per million tokens (MTok), with separate rates for input, cached input, and output. Most providers copy this structure and then apply their own margin on top. That's why the number on your invoice can differ from the number on DeepSeek's site — the model is the same, the route is different.
Official vs. third-party pricing
Calling DeepSeek directly gives you the raw rate, but you also take on the operational work: key management, rate limits, per-region latency, and monitoring. Gateways bundle that work into their price.
Some providers charge a premium for convenience; others run thin margins and make money on volume. A few also pass through cache discounts properly, which matters a lot for production workloads with long system prompts.
Why cache hits change the math
If your traffic has a high prefix-cache hit rate, the effective cost per token can drop by 80-90%. Not every reseller exposes cache-hit pricing though, so two providers quoting similar list prices can produce very different real bills. Always compare the cache-hit rate, not just the headline number.
Where to compare DeepSeek API pricing side by side
I keep a live comparison table that updates automatically whenever upstream prices change. If you want to compare DeepSeek API pricing across providers in one place, you can check the KeyoAPI DeepSeek rates page — it lists direct and gateway pricing, cache-hit rates, and context-window details for the current models.
Quick ways to cut your LLM API bill
- Route repetitive tasks through cached prompts so you earn cache-hit rates.
- Batch non-urgent jobs to lower-bandwidth endpoints.
- Use a smaller model for classification and routing, and keep the frontier model only for hard problems.
- Check your provider's actual per-token cost each month — list price and effective price drift apart.
DeepSeek pricing is changing fast. Whatever you build, the important thing is to measure your real cost per million tokens on the route you actually use, not the one you assumed.
Top comments (0)