DeepSeek API Pricing 2026: Real Cost With Top-Up Fees

Last updated 2026-09-17 · Written by Nova

Every pricing page shows a per-million-token number. Almost none of them show what you actually pay once the payment rail takes its cut and the minimum top-up forces you to park money you didn't plan to spend.

The token price is the easy part

Rates here, per million tokens, for the models we serve:

How that compares with the models you'd otherwise be using

The comparison below is deliberately relative: we publish the saving against Claude Sonnet's public list rates rather than a token rate, because the percentage is what shows up on a monthly bill and the rate is not.

Read the first two rows against the rest: on deepseek-flash that is roughly 70–85% less, on input and output alike, for work where a coding agent or a chat tool will not notice the difference. The endpoint speaks both the OpenAI and the Anthropic format, so moving across is a base-URL change rather than a rewrite.

Two things we would rather you hear from us than discover later. Our prices are not the lowest numbers quoted in this market — some gateways advertise below the model vendor's own rate, which usually means the supply behind them is borrowed rather than bought, and it tends not to last. And peak hours cost more: DeepSeek applies automatic peak pricing during Chinese business hours and our rates follow that mechanism, so budget for it if your traffic runs 09:00–12:00 or 14:00–18:00 Beijing time on a weekday.

Those are the numbers people compare and they're the least interesting part of your bill. A typical chat workload is roughly 2 parts input to 1 part output, so a blended cost lands around $0.9 per million tokens on deepseek-flash.

The four costs nobody puts on the pricing page

1. The minimum top-up. A provider quoting cheaper per-token rates but requiring a $50 minimum is more expensive than a $5 minimum when you're testing an idea. You're not buying tokens, you're parking cash in someone's dashboard. We start at $5.

2. The payment rail's cut. This is the big invisible one for anyone outside the US and EU:

• PayPal takes a currency conversion spread, commonly around 4% on top of the rate. On $5 that's roughly $0.20 gone before you've bought a single token; on $10 it's about $0.40.
• USDT on TRC20 costs about $1 flat, regardless of size — so ~20% on a $5 top-up, ~10% on a $10 one, and ~1% on a $100 one. Larger, less frequent top-ups are meaningfully cheaper.
• Buying the USDT itself costs something too, via P2P spread. Usually under 1%, but not zero.

3. Failed renewals. Auto-renewing card subscriptions fail silently. Your first month clears, the second doesn't, and your app is down at 3am with no alert. A prepaid balance doesn't have this failure mode at all — which is worth more than a few percent on price.

4. Your own time. Chasing a declined card, opening a support ticket, migrating to a new provider mid-project. That's the most expensive item on this list and it never appears in a comparison table.

A worked example

Say you're running 10 million tokens a month, two-thirds input:

The same month paid by PayPal carries roughly 4% more in conversion spread; the same month on a provider with a $50 minimum means you are holding a much larger balance before you can spend anything at all.

How to spend less without switching providers

• Send less context. Most waste isn't output tokens, it's re-sending the same system prompt and history on every turn. Cost grows roughly quadratically with conversation length if you resend everything.
• Top up less often, in bigger chunks. The TRC20 fee is flat, so $50 once beats $10 five times.
• Use the small model for the boring calls. Routing classification and extraction to deepseek-flash and reserving the pro model for hard reasoning is usually a bigger saving than any provider switch.
• Watch the first month. Whatever you modelled, the real number is what your workload does — measure it before optimising.

Honest disclosure

We're a reseller, not DeepSeek. How the price compares: for the same workload, these rates land about 70–85% below Claude Sonnet’s list price. We are not the cheapest way to buy tokens, and we won't pretend otherwise. We're worth it when the payment rail is your actual blocker, not when price is your only criterion.

One production node in Hong Kong. No SLA, and we publish no availability percentage rather than a figure we have not measured. We don't store prompt contents beyond billing and abuse prevention.

Frequently asked questions

Are the token prices peak and off-peak?

Yes. DeepSeek applies automatic peak pricing during Chinese business hours, which doubles the per-token cost for those hours and returns to normal outside them and throughout the weekend. We follow that mechanism as-is; it is not our invention and it is not something we can switch off.

How much does a typical month actually cost?

For around 10 million tokens a month at roughly a two-to-one input-to-output ratio, the token subtotal works out close to $9, and the payment rail adds about $1 for a USDT transfer, or roughly 4% with PayPal. Treat that as an illustration of method rather than a quote for your workload — real usage sends the same system prompt and history on every turn, which pushes the effective total above a naive token count.

What is the cheapest way to reduce the bill?

Send less context, because the waste is usually re-sending the same system prompt and history on every call rather than generating output. Top up less often in larger amounts, since a flat transfer fee is charged per transfer. Route routine calls such as classification and extraction to the smaller model and reserve the larger one for harder tasks.

Is this cheaper than going direct to DeepSeek?

No — our rates include a margin, so if your card works and you can pay the official platform directly, direct is cheaper per token. A reseller makes sense when the payment channel is the problem, not when price is. That is the honest comparison and we would rather you make the decision with it.