Know where your DeepSeek balance is heading

How does DeepSeek bill for usage?

DeepSeek runs on a prepaid balance. You add credit to your account, and every API call deducts from it as you go — there’s no monthly invoice and no subscription tier to reason about, just a number that goes down. DeepSeek draws from two pools when you have them: a granted balance (promotional or complimentary credit) is spent first, and your own topped-up balance covers the rest. Both drain the same way, and the account view shows their sum.

What pulls that balance down is metered the way most token APIs are: by the input tokens you send — your prompt, system instructions, and any context — and the output tokens the model generates back. Output is the pricier side, so a long or chatty completion costs more than the size of the prompt suggests.

DeepSeek splits into two modes, and the difference matters for cost. There’s a plain non-thinking mode for general work, and a thinking mode that reasons through a problem before it answers. That reasoning isn’t free. The model produces a chain of thought first, and those tokens count toward the output length you’re billed for, even though they’re working notes and not the final reply. A request that looks small can carry a large, mostly invisible output charge when thinking mode is doing real work. That’s the part that catches people out the first time they lean on it. (DeepSeek is mid-transition on model names — the long-standing deepseek-chat and deepseek-reasoner are scheduled to be retired in mid-2026 in favor of the deepseek-v4 lineup, deepseek-v4-flash and deepseek-v4-pro — so treat any specific model name as a moving target and check the docs for the current set.)

One mechanism works in your favor. DeepSeek prices repeated prompt prefixes through context caching: when a later call reuses input the service has already seen, that portion bills at a much cheaper cache-hit rate instead of the full cache-miss rate. Keep your stable prefixes — long system prompts, reference material — consistent across calls and a real slice of your input cost moves to the cheaper tier.

So a single DeepSeek call can draw on three token prices at once, each quoted per 1M tokens:

Token type	How DeepSeek prices it
Input — cache miss	The full input rate, for prompt content it hasn’t seen
Input — cache hit	A fraction of that rate, when a later call reuses a prefix
Output	The expensive side — what the model generates back. Reasoning-mode thinking tokens are counted here too, at the same output rate

Exact rates, quoted per 1M tokens, live on DeepSeek’s pricing page and have changed before, so treat any figure you’ve memorized as provisional.

Why is DeepSeek spend hard to keep ahead of?

The pricing itself is simple. Staying ahead of it is the hard part, and most of the difficulty comes down to what DeepSeek does and doesn’t show you:

• The account balance is the only signal to poll. DeepSeek’s account API hands back your current balance and nothing else: no account-level usage history, no per-model report to break down or replay. You can see where you stand, but there’s no account endpoint to ask where the money went.
• Prepaid hides the slope. A balance tells you the level. It says nothing about the rate it’s falling at. By the time the number looks low enough to notice, the spend that pulled it down is already behind you, and a countdown gives no early warning.
• Thinking tokens are invisible. A reasoning-heavy request can burn far more output than its prompt implies, and the chain of thought that drove the charge isn’t part of the final answer. From the account balance, it shows up only after the fact.
• DeepSeek is rarely your only provider. If you also call Anthropic for Claude, build on OpenAI, run Gemini, or route through OpenRouter — plus whatever you spend on inference hosts and serving — the DeepSeek balance is one slice of a bill you otherwise stitch together by hand from a stack of dashboards.

How can you reduce your DeepSeek bill?

Tracking shows you where the balance goes. A few levers slow the drain:

• Match the mode to the task. Thinking mode earns its keep on hard problems, but it bills the reasoning tokens as output. Route the straightforward, high-volume calls to the non-thinking mode and save the reasoning model for work that needs it. That one routing choice tends to move the bill more than any other tweak.
• Lean on context caching. Keep large, stable prefixes consistent so they land on the cheaper cache-hit rate instead of paying full input on every call. Long system prompts and reference documents are where this pays off.
• Keep outputs tight. Output is the expensive side, and thinking tokens ride along with it. Trimming verbose responses and reining in how much the model reasons on tasks that don’t need deep thought cuts the part of the bill that grows fastest.
• Watch the trend, not the balance. A number that only goes down tells you little until it’s nearly gone. A running total with a slope on it shows a creeping spike within a few days, well before the account runs dry.

How do you forecast next month’s DeepSeek bill?

Forecasting a prepaid balance doesn’t need anything exotic, just a burn rate: take your spend over the last several days, turn it into a daily average, and project that across a full month.

CostCompass does exactly this. It scales your trailing seven-day burn rate to the number of days in next month and adds any fixed subscriptions you’ve entered. The result is one forward number for what next month costs if the current pace holds. Seven days is a deliberate window: long enough that one heavy day doesn’t throw the average, short enough to pick up a recent change like a shift to thinking mode. For DeepSeek the projection builds on a balance-derived total, so read it as a directional estimate. It gets the order of magnitude right and sharpens as more clean windows land, though it won’t be pinned to the last cent.

How does CostCompass track your DeepSeek costs?

CostCompass tracks DeepSeek by watching the one thing DeepSeek will tell it: your account balance. Each time you refresh, it reads the current balance and compares it against the previous reading, and the drop between the two becomes the spend for that window. There’s no account-level usage report behind this for CostCompass to read, so the figure is the total you spent across DeepSeek, marked approximate with a small asterisk. It can’t be a per-model breakdown, because the data to support one doesn’t exist. Tracking begins with your first refresh after connecting, which captures a baseline and shows no spend yet; from the next refresh on, every dip in the balance rolls into your running total.

That approach has two edges worth knowing up front. Spend from before your first refresh isn’t recoverable, since there’s no history to read back, so the total starts from that first reading instead of the first of the month. And a top-up in the middle of a window can mask the spend it overlaps. If you add credit and burn some in the same stretch, the balance may end higher than it started, so that window under-counts. A later window starts clean from the new balance and won’t recover the masked spend. Neither edge is something you can act on — they’re just what derived spend looks like — but it’s worth knowing the masked stretch reads low and never fills back in.

Two things make that practical for a solo developer. First, your key is encrypted in your browser before anything is stored. The DeepSeek key you paste is sealed with your vault password on your own device, and only the sealed version is ever saved — a blob CostCompass can’t open, because the vault password stays with you. When it’s time to read your balance, your browser unseals the key and passes it to a relay that uses it for that single call to DeepSeek and is built not to log or keep it. What we hold at rest is locked ciphertext that nobody can use against your account.

Second, DeepSeek doesn’t sit alone. The same dashboard rolls its running total up with every other AI and compute provider you’ve connected — Claude, OpenAI, Gemini, your hosting — into one figure with a forecast. DeepSeek becomes a line in your whole-stack spend instead of a balance you check in a console of its own.

Getting started takes three steps:

1. In DeepSeek, create an API key — the ordinary kind, no special permission needed.
2. Paste it into CostCompass. It’s encrypted in your browser before it’s stored, so the server only ever holds ciphertext.
3. CostCompass captures a baseline balance on your first Refresh, and from the next Refresh on it shows your running total and forecast — DeepSeek included in the whole-stack number.