DeepSeek API pricing: the value king, by the numbers

Quick answer: what the DeepSeek API costs in 2026

DeepSeek V4 Flash bills $0.126 per million input tokens and $0.252 per million output tokens. DeepSeek V4 Pro bills $0.3915 in and $0.783 out. Those are the two current models, with IDs deepseek-v4-flash and deepseek-v4-pro, and the gap between them and Western flagships is not a rounding error. Put it in request terms. A typical chat turn with 2,500 tokens of accumulated context and a 300-token reply costs about $0.0004 on Flash, which works out to roughly 2,500 turns per dollar. The same turn on Pro costs about $0.0012, around 800 turns per dollar. On Claude Sonnet 4.6 that turn costs about $0.0096, and on gpt-5.5 about $0.017. If you take one thing from this page: for chat, roleplay, and agent loops, DeepSeek's cost per interaction sits one to two orders of magnitude below flagship pricing, and in 2026 the quality gap is far smaller than that price gap.

Why DeepSeek is the value king: one workload, six price tags

Per-token rates are abstract, so here is the same monthly workload priced across current models: 10 million input tokens and 2 million output tokens, roughly what a busy hobby project or a small production assistant consumes in a month. The pattern in the table repeats at any scale. DeepSeek V4 Flash is the cheapest capable model in the catalog by a wide margin, V4 Pro undercuts every Western flagship while competing on reasoning tasks, and GLM-5 is the nearest value alternative. A datapoint that circulates in agent and roleplay communities: one user reported pushing about 28 million tokens through DeepSeek in a month for roughly $2.82. Treat that as an approximate, self-reported figure rather than a guaranteed rate; numbers that low usually involve prompt caching on the direct platform discounting repeated context. But you do not need the best case to make the point. At plain catalog rates, 28 million tokens on V4 Flash lands between about $3.53 and $7.06 depending on your input and output split. The same volume on Claude Sonnet 4.6 would run between $67 and $336, and on gpt-5.5 between $112 and $672. DeepSeek turns the monthly LLM bill from a budget line into pocket change.

DeepSeek V4 Flash vs V4 Pro: which one do you actually need?

Default to Flash and escalate only when you can name the failure. Flash handles conversation, extraction, summarization, and straightforward agent steps at the lowest cost per token in its class. Pro costs about three times as much per token and buys you deeper multi-step reasoning, better performance on hard code, and more coherent handling of long, tangled instructions. The escalation test is simple: run your worst real prompt on Flash five times. If it holds up, you just cut your bill by two thirds. If it drops instructions or loses the plot mid-task, move that one workload to Pro and keep everything else on Flash. Because both models sit behind the same API shape, routing between them is a one-string change.

What people actually run on DeepSeek

Where does DeepSeek stop being the answer? If your product depends on the best available prose for safe-for-work creative writing, Claude Sonnet 4.6 remains the quality ceiling, at a matching premium, and Anthropic's policy restricts creative work to non-explicit content. Kimi K2.6 and MiMo are worth benchmarking as additional value picks for chat-heavy loads. For everything measured in tokens per dollar, DeepSeek is the default in 2026.

• Agent loops: agents re-send a growing context on every step, so input price dominates the bill. A fifty-step run with an average 8,000-token context costs about five cents on Flash before output.
• Roleplay and character chat: long chats replay the full history each turn, which makes per-token price the deciding factor. The DeepSeek V4 family is the community favorite on SillyTavern and JanitorAI for exactly this reason, and it handles long-form fiction with a comparatively flexible content policy for fictional scenarios. Keep scenarios within each platform's rules and age settings.
• Batch processing: classification, dataset labeling, summarizing document piles, and synthetic data generation are throughput problems. At Flash rates, a million short classifications cost less than a takeout lunch.
• Prototyping: when you are iterating on prompts all day, a cheap model changes behavior. You stop rationing test runs and start actually testing.

How to access the DeepSeek API (direct or through a gateway)

There are two practical routes. Route one is DeepSeek's own platform: create an account, set up billing, and call their endpoint. Rates are excellent, and it is the right choice if DeepSeek is the only model you will ever call. Route two is an OpenAI-compatible gateway such as apisrouter.com, which puts deepseek-v4-flash and deepseek-v4-pro next to glm-5, kimi-k2.6, claude-sonnet-4-6, gemini-3.5-flash, and gpt-5.5 behind one base URL (https://api.apisrouter.com/v1) and one pay-as-you-go balance, so mixing or switching models is a one-line change. There is no subscription and no account form: pay at /topup, the key arrives by email, the first top-up doubles your balance, and catalog pricing runs below official rates for Chinese models and 20% off official for global models. Front ends follow the same pattern. SillyTavern takes the base URL ending at /v1 and auto-populates the model list from /v1/models once your key is in. JanitorAI's proxy config wants the full endpoint, https://api.apisrouter.com/v1/chat/completions, plus the model name and API key. The first request is one curl away:

curl https://api.apisrouter.com/v1/chat/completions \
  -H "Authorization: Bearer sk-..." \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-v4-flash",
    "messages": [
      {"role": "user", "content": "Summarize this support ticket in two lines: ..."}
    ]
  }'

Track your real cost per request

Every response includes exact token usage, so cost tracking is three lines of arithmetic rather than a dashboard subscription. This is the fastest way to settle the Flash vs Pro question for your own workload: log cost per request for a day on each and compare failure rates against the price delta.

from openai import OpenAI

client = OpenAI(
    api_key="sk-...",
    base_url="https://api.apisrouter.com/v1",
)

resp = client.chat.completions.create(
    model="deepseek-v4-pro",
    messages=[{"role": "user", "content": "Plan the next three agent steps for: ..."}],
)
print(resp.choices[0].message.content)

# V4 Pro catalog rates: $0.3915 in, $0.783 out, per 1M tokens
usage = resp.usage
cost = usage.prompt_tokens * 0.3915 / 1e6 + usage.completion_tokens * 0.783 / 1e6
print(f"request cost: ${cost:.6f}")

FAQ