DeepSeek Billing and API Pricing 2026: Real Cost Per Run
DeepSeek billing is per million tokens, with peak and off-peak rates and cache hit or miss pricing. Here is the current rate card, how balance is deducted, and how to cut agent spend.

DeepSeek billing is usage based. Each request is charged as tokens multiplied by the applicable rate per 1 million tokens, with input and output priced separately, and fees come out of your topped-up or granted balance, granted balance first. Rates then split by peak versus off-peak and by cache hit versus cache miss, so one model carries six prices, not one. Checked on 6 November 2026 against DeepSeek's official pricing page, off-peak deepseek-flash is $0.15 per million cache-miss input tokens and $0.60 output.
The rate card is only half of what you pay. In an agent, the larger half is how many times the model re-reads the same context and re-derives the same step. Major, the enterprise platform where agents build the software they run on, attacks that term: the agent reasons once, builds an app for the repeatable part, and runs the app from then on. The section on building it comes after the numbers.
How DeepSeek billing works
Charges equal token usage multiplied by the rate for the model and tier. Fees are deducted from the available topped-up or granted balance, and when both exist the granted balance is used first. Search and Fetch tools are billed separately, per successful request, at $0.75 per 1,000 requests each.
Two practical consequences. Credits you were granted burn before money you topped up, so your first invoice-equivalent figure understates steady-state cost. And tool calls appear as their own line, so an agent that searches on every step has a second meter running next to the token meter.
DeepSeek API pricing: the current rate card
All prices are USD per 1M tokens, taken from DeepSeek's official docs on 6 November 2026. Both models have a 1M-token context window.
deepseek-flash(DeepSeek-V4.1-Flash-1017), off-peak: $0.003 cache-hit input, $0.15 cache-miss input, $0.60 output.deepseek-flash, peak: $0.006 cache-hit input, $0.30 cache-miss input, $1.20 output.deepseek-v4-pro(DeepSeek-V4-Pro-0813), off-peak: $0.022 cache-hit input, $0.66 cache-miss input, $1.98 output.deepseek-v4-pro, peak: $0.044 cache-hit input, $1.32 cache-miss input, $3.96 output.
The retired names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted. DeepSeek states that prices may change and the page carries no changelog, so stamp the date you checked and confirm against the source before budgeting. Older rate cards, and many third-party pages, still list the previous flash prices.
DeepSeek peak and off-peak hours
Off-peak is half of peak. Peak runs 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, except Chinese public holidays. Everything else, including weekends and those holidays, bills at the lower tier.
Schedule anything batchable outside those seven weekday hours: a nightly enrichment run, a classification pass, a weekly report build. Same tokens, half the price. Interactive traffic follows your users' clock, so budget it at peak rates and treat off-peak as upside. A London team working business hours buys peak tokens every morning without having chosen to.
Cache hit versus cache miss: the biggest lever
On deepseek-flash off-peak, cache-miss input is $0.15 per million and cache-hit input is $0.003, a 50x gap on identical tokens. On deepseek-v4-pro the gap is 30x.
Cache hits reward a stable prompt prefix. Keep the system prompt, tool definitions and retrieved context in the same order at the front of every request. Put a timestamp at the top, or shuffle retrieved documents per call, and the whole prompt bills at cache-miss rates indefinitely. Prompt layout is a billing decision.
What DeepSeek costs at your volume
Every input below is an assumption. Swap in your own; the arithmetic holds.
- Low-volume internal feature. 2,000 calls a month, 3,000 uncached input tokens and 500 output tokens per call,
deepseek-flash, off-peak. Input is 6M tokens at $0.15, so $0.90. Output is 1M at $0.60. Total $1.50 a month, $3.00 at peak. Pick on quality and stop reading pricing pages. - High-volume classification. 5M documents a month, 800 input tokens each with a 600-token stable prefix, 20 output tokens,
deepseek-flash, off-peak. Cache-hit input is 3,000M tokens at $0.003 ($9.00), cache-miss input is 1,000M at $0.15 ($150.00), output is 100M at $0.60 ($60.00). Total $219.00. With an unstable prompt nothing caches, input becomes 4,000M at $0.15, and the bill is $660.00. Same model, three times the cost. - Agent loop that re-reads context. 6,000 runs a month, six calls per run, a 40,000-token context re-read on every call, 700 output tokens per call,
deepseek-v4-pro, off-peak, first call a miss and the other five hits. Cache-miss input is 240M at $0.66 ($158.40), cache-hit input is 1,200M at $0.022 ($26.40), output is 25.2M at $1.98 ($49.90). Total $234.70, or $0.0391 per run.
That last figure is cost per attempt. If 80% of runs succeed first time and the rest need one full retry, 6,000 completed tasks consume 7,200 attempts. The bill becomes $281.64 and cost per completed task is $0.047. Retries added 20% from a factor that appears nowhere on the rate card.
One constraint can outrank price at volume: account-level concurrency limits, with over-capacity requests returning HTTP 429. Check your limits in DeepSeek's docs before you size a batch job.
Why third-party DeepSeek pricing pages disagree
Four dimensions give aggregators many numbers per model to get wrong, and DeepSeek has renamed and repriced its models more than once. A page that quotes one flat input price has collapsed peak, off-peak, hit and miss into a single figure. Source DeepSeek prices from DeepSeek. For model-by-model comparisons across vendors, see LLM API pricing: the rate card is not the bill and the cheapest LLM API and why the lowest token price rarely wins.
Build the cost-controlled DeepSeek agent on Major
In the agent-loop scenario, $184.80 of the $234.70 is re-read context. A cheaper model lowers the price of a unit of reasoning. It does not change how many units the agent buys. Major moves that term, and it is the best way to run a DeepSeek-backed agent whose spend stops climbing with usage. Start building free and try it on one workflow.
- Agent: calls the DeepSeek model for judgment only, such as classifying an ambiguous document or choosing a route. When it works out a repeatable step, it builds an app for it.
- App layer: a Major app holds the deterministic steps as code, with a managed database for state, storage for files and its own logs. The context the model used to re-read persists in the app.
- Connectors: pull inputs from your system of record, such as Snowflake or BigQuery, and post exceptions to Slack for a person to review.
- Workflow trigger: a schedule that fires batch jobs outside the 01:00-04:00 and 06:00-10:00 UTC weekday peak window.
- Governance: scoped credentials per connector, role-based access to the app and an audit log of every action the agent takes.
Cost is front-loaded then flat. Working out the workflow costs real reasoning once, and repeat runs do not scale token spend with usage. The model does less, not nothing. Reason once. Run forever. For the wider playbook, read LLM cost management: control spend by running less AI and AI model selection: choose per task, not by hype.
Other model rate cards
If you are budgeting across vendors, the same four-way split of tier, cache state, input and output applies elsewhere. See Qwen API pricing and Gemini API pricing for their current cards, then compare on cost per completed task rather than the headline rate.
What this article doesn't cover
Fine-tuning and batch-mode pricing, non-USD billing, reseller rates for open DeepSeek weights, and latency under load.
Start building on Major
Build an agent that turns its repeatable steps into governed apps, so your DeepSeek spend stops climbing with every run.
Related articles
Frequently asked questions
- How does DeepSeek billing work?
- DeepSeek charges token usage multiplied by the rate for the model and tier, per million tokens, with input and output priced separately. Fees are deducted from your topped-up or granted balance, and granted balance is used first. Search and Fetch tools are billed per successful request at $0.75 per 1,000 requests each.
- How much does the DeepSeek API cost?
- Checked 6 November 2026, deepseek-flash costs $0.003 to $0.30 per million input tokens and $0.60 to $1.20 output. deepseek-v4-pro costs $0.022 to $1.32 input and $1.98 to $3.96 output. The spread comes from cache state and peak versus off-peak hours.
- What are DeepSeek peak hours?
- Peak is 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, except Chinese public holidays. All other times, including weekends and those holidays, are off-peak and billed at half the peak rate. Schedule batch jobs off-peak to halve their token cost.
- Is DeepSeek free?
- The consumer chat product is free to use, but the API is paid. Charges are deducted from a topped-up or granted balance as tokens are consumed, with granted balance used first. Any free credit you receive is a balance that runs out, not an unlimited allowance.
- Why do DeepSeek pricing pages show different prices?
- Each model has peak and off-peak rates plus cache-hit and cache-miss input prices, so aggregators often collapse six numbers into one, and DeepSeek has renamed and repriced models. Use DeepSeek's official pricing page and note the date you checked.
- Is a cheaper model actually cheaper to run?
- Only if it needs the same number of attempts. Workload cost is tokens per attempt, times rate per token, times attempts per completed task. A cheaper model that needs retries or a fresh context read on every call can cost more in production. Track cost per completed task.