DeepSeek API Gateway Pricing: What Will You Actually Pay?

A low price per million tokens is not enough information to choose a DeepSeek API gateway.

You still need to know which tokens that price applies to, whether cached input is billed separately, which model your request reaches, and how the amount becomes a deduction from your account.

For a developer comparing API relay services, the useful question is simple:

Can I calculate the cost of a request before sending it, then explain the charge afterward?

This guide shows how to answer that question without relying on a headline discount.

Start with three prices, not one

A DeepSeek model request can contain three relevant billing quantities:

The official DeepSeek pricing page lists these categories separately.

For reference, the published direct-service rates checked on November 4, 2026 were:

All figures are USD per one million tokens. The official page also lists off-peak rates and the applicable schedule.

These are direct DeepSeek reference prices. They are not a quotation for an intermediary's account, group, or channel.

A gateway may use its own billing arrangement. Compare its actual terms with the reference rather than assuming that every upstream pricing rule passes through unchanged.

Work through one request

Consider a hypothetical DeepSeek Flash request:

• Total input: 100,000 tokens.
• Cached input: 20,000 tokens.
• Uncached input: 80,000 tokens.
• Output: 10,000 tokens.

Using the peak direct-service reference above:

Uncached input:
80,000 ÷ 1,000,000 × $0.30 = $0.024

Cached input:
20,000 ÷ 1,000,000 × $0.006 = $0.00012

Output:
10,000 ÷ 1,000,000 × $1.20 = $0.012

Model-token subtotal:
$0.024 + $0.00012 + $0.012 = $0.03612

This is a worked calculation, not a measured request.

One thousand identical requests would produce a model-token subtotal of $36.12 under those assumptions. Real workloads will vary in input length, output length, cache behavior, and applicable pricing.

Notice the accounting detail: cached input is part of total input. Do not bill all 100,000 input tokens as uncached and then add another charge for the same 20,000 cached tokens.

Also use the service's billed usage categories rather than estimating output from the visible answer alone.

Translate the gateway quote into the same units

Before comparing providers, normalize the offer.

Write down:

Exact model:
Account group or route:
Billing currency:
Uncached input price per 1M:
Cached input price per 1M:
Output price per 1M:
Other applicable charges:
Effective date:

If the service quotes account credits, obtain the conversion between credits and payable currency. If it quotes a channel adjustment, determine whether that adjustment is already included in the displayed unit prices.

Applying the same adjustment twice can make your spreadsheet wrong even when the platform's charge is correct.

Likewise, an account balance displayed in dollars is not enough to establish how a local-currency payment was converted.

For TreeRouter, review the pricing and account-group documentation together with the group selected for your token. Use the current account quote for your calculation rather than a price copied from an old article.

Cache availability and cache savings are different questions

DeepSeek documents context caching, but a gateway evaluation requires two separate checks:

1. Does the request qualify for upstream cache reuse?
2. How does the gateway report and bill the resulting usage?

The official context-caching guide explains prefix reuse and cache-hit behavior.

Repeating a short prompt does not establish a dependable cache-hit ratio for your production workload. Nor does an absent cache field prove that no cache was used.

If a response exposes only total input usage, use the billing record or obtain a clear explanation from the provider. Do not invent a cache split to make the calculation match.

For early budgeting, calculate a conservative case with no cache savings, then replace the assumption with observations from representative traffic.

Reconcile one charge before estimating a month

A useful first test is deliberately small.

Create a dedicated test token, send a bounded request, and save:

• The request time.
• Requested model and selected account group.
• Response identifier, when available.
• Returned usage fields.
• The corresponding account log entry.
• The displayed charge and currency.

TreeRouter's usage-log guide explains how to expand a log entry to inspect token consumption and its calculation.

Match the response to the relevant log. If the platform does not expose a shared request identifier, use an isolated test token and a narrow time window; avoid sending concurrent test requests that make matching ambiguous.

Small differences may involve rounding or billing units. Large unexplained differences need investigation before you increase traffic.

Compare usable results, not just successful responses

Suppose one route is cheaper per token but repeatedly truncates the output your application needs.

Its lower unit price may not translate into a lower cost per accepted task.

For a representative test set, calculate:

Cost per accepted task =
total billed cost for the test set
÷ number of outputs accepted by your application

Include billed retries and rejected outputs in the numerator. Keep the acceptance rules identical across candidates.

This is not an argument for buying the most expensive route. It is a way to distinguish a low price from a low operating cost.

The related model acceptance guide provides a starting point for defining those rules.

Keep model charges separate from auxiliary services

A workflow may also use search, page fetching, storage, or other tools.

Those services do not necessarily share the model's token-based billing unit. Keep them on separate lines in your estimate.

A gateway supporting DeepSeek text generation does not automatically include every adjacent DeepSeek service in the same price.

This distinction matters most when comparing a plain chat request with an agent task that performs several external operations.

Make your first purchase answer a concrete question

If you are evaluating TreeRouter, create an account, select the intended DeepSeek route, and run a small test appropriate to the account's billing options.

The first milestone is not a large balance. It is a charge you can explain.

Once one request reconciles, test a representative workload and compare the actual cost per accepted task. That gives you a defensible reason to expand usage.