DeepSeek API Pricing 2026: Official Rates per Million Tokens

Rates verified 2026-08-18. Every DeepSeek number on this page was read off DeepSeek’s own pricing page on that date, after the peak/off-peak change that took effect at 16:00 UTC on 2026-08-16. For how much each tier moved and why, see DeepSeek API Price Increase.

TL;DR: DeepSeek V4 Flash costs $0.22 per million input tokens off-peak and $0.44 at peak, with output at $0.66 and $1.32. V4 Pro is three times that on input and output, and a shade more than three on cache hits. A cache hit costs about 3% of a cache miss, which is the largest single lever on the bill. Peak is seven hours a day, 01:00 to 04:00 and 06:00 to 10:00 UTC, and off-peak is exactly half price. If you would rather have one key across DeepSeek, Claude, GPT and Gemini than chase a clock, ofox.ai carries both models at the peak rate around the clock.

Try DeepSeek V4 via ofox.ai with one OpenAI-compatible key.

What Is DeepSeek V4 Flash?

DeepSeek V4 Flash is DeepSeek’s current volume model, offered alongside the more capable V4 Pro. Both support a 1M token context window and generate up to 384K output tokens in a single response. On the API today, deepseek-v4-flash resolves to DeepSeek-V4-Flash-0731 and deepseek-v4-pro resolves to DeepSeek-V4-Pro-0813.

V4 Pro adds deeper reasoning for multi-step problems, which is what you pay three times the token price for. V4 Flash handles the ordinary work, summarization, classification, chat and routine coding, and both models default to thinking mode with a documented way to switch it off.

Two IDs that used to be in every DeepSeek tutorial, deepseek-chat and deepseek-reasoner, are no longer listed in the model table on the pricing page or on the first-call page as of 2026-08-18. If your code still sends them, that is worth testing rather than assuming.

What Is the Official DeepSeek API Pricing?

All figures are per million tokens, from DeepSeek’s pricing page on 2026-08-18. Off-peak is exactly half of peak on every line.

Peak is 01:00 to 04:00 and 06:00 to 10:00 UTC, seven hours a day. The other seventeen hours are off-peak. That is the whole mechanism: there is no tier to opt into, no flag to set, and no way to buy the off-peak rate during peak hours.

The cache line is where the real money is. A hit costs about 3% of a miss on the same tier, roughly a 97% discount on that token, and it applies automatically whenever a prefix of your prompt matches something DeepSeek already has. Nothing to configure and no separate endpoint. The reason the cache column is worth staring at is that most production prompts are mostly prefix: a system prompt, a tool schema, a document that does not change between turns.

One thing DeepSeek does not document: whether a request that starts off-peak and finishes at peak is billed by its start time or its end time. If you run long sessions near 04:00 or 10:00 UTC, that is an open question rather than a known answer.

How Much Does a Real Month Cost?

Take a workload of 10M input tokens and 2M output tokens a month on V4 Flash, and price it three ways.

The gap between the first two rows is the argument for structuring prompts around a stable prefix. The gap between the two columns is the argument for moving anything batch-shaped, such as document processing or dataset annotation, out of the seven peak hours.

Put that in per-request terms. A 2,000 token system prompt that gets cached on every call saves $426 per million requests off-peak, and $852 per million at peak, versus paying cache-miss rates for the same tokens. That is the single largest line item you control without changing models.

How Do You Get a DeepSeek API Key?

Sign up at platform.deepseek.com with an email address. No Chinese phone number is required for the API platform, unlike the consumer app. Once logged in, go to API Keys and generate a key.

The base URL is https://api.deepseek.com for the OpenAI-compatible route and https://api.deepseek.com/anthropic for the Anthropic-compatible one, so the same key works from either SDK by changing base_url. JSON output, tool calls, the Responses API and chat prefix completion are all supported on both models; FIM completion is beta and non-thinking mode only.

Billing runs off a topped-up balance and a granted balance, with the granted balance spent first when you have both. There is no documented recurring free tier, so treat any signup credit as a trial rather than a plan.

One practical note: DeepSeek publishes a concurrency cap rather than an uptime commitment, and there is no public status page to point at. If you are building something production-critical, plan the fallback yourself rather than assuming headroom.

How Do You Access DeepSeek via ofox?

ofox.ai carries DeepSeek V4 Flash as deepseek/deepseek-v4-flash-0731 at $0.44 input and $1.32 output, and DeepSeek V4 Pro as deepseek/deepseek-v4-pro-0813 at $1.32 input and $3.96 output. Those are the official peak rates, charged around the clock, with no peak/off-peak split.

Say that plainly, because it cuts both ways. During the seventeen off-peak hours, going direct to DeepSeek costs half what the gateway costs. During the seven peak hours the two are the same price. What the gateway buys you is one key across every model you use, one bill, and somewhere to fail over to when the official platform is busy. If DeepSeek is the only model you call and your jobs are batch-shaped, direct is cheaper and you should say so out loud in your own cost review.

One detail worth knowing if you care about caching: the same model ID can route through more than one upstream, and the cache-read rate is not the same on all of them. On the DeepSeek and BaiLian routes, cache reads on V4 Flash bill at $0.014 per million; on the Azure route they bill at $0.44, the full input price, which means no cache discount at all. The live per-route breakdown is on the DeepSeek model page.

The practical advantage is a single API key that also covers Claude, GPT, Gemini, Qwen and the rest of a catalogue that listed more than 130 models on 2026-08-18. No separate accounts, no separate billing. This matters if you route between models by task type or cost, something covered in more depth in the AI API aggregation guide.

Setup is three lines:

from openai import OpenAI

client = OpenAI(api_key="sk-xxx", base_url="https://api.ofox.ai/v1")
response = client.chat.completions.create(
    model="deepseek/deepseek-v4-flash-0731",
    messages=[{"role": "user", "content": "Explain the 1M token context window in one paragraph"}]
)

If you are migrating from the OpenAI SDK, the migration guide covers the full swap in under 10 minutes.

How Do You Cut Your DeepSeek API Costs?

The cache discount is the biggest lever available, and it is bigger than it was: when input was $0.14 a cache hit saved you $0.137 per million tokens, and at $0.22 off-peak it saves $0.213. Structure prompts so the system prompt and any static context come first, because those are the parts most likely to be cached. You do not configure anything; the API applies it when a prefix matches. The cache hit math walkthrough covers why real bills usually show a lower hit rate than the design implies.

The second lever is new. Move anything that does not need to run now out of the peak window. Nightly document processing, dataset annotation and bulk summarization are all schedulable, and off-peak is half price on every line including cache reads. The seven peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, so what that costs you in convenience depends entirely on where you are; teams in the Americas hit almost none of it during a normal workday, and teams in East Asia hit most of it.

The third is model selection. Only use deepseek-v4-pro when you actually need deeper reasoning. Pro costs exactly three times Flash on cache-miss input and on output, so classification, summarization and simple generation on Pro is a 3x bill for no gain.

Context trimming still matters. The 1M window is large, but you pay for every input token that is not a cache hit. RAG that pulls only the relevant chunks is cheaper than stuffing the full document, and the embedding and RAG guide covers that pattern. For a broader playbook across all models, see how to reduce AI API costs.

DeepSeek vs. Alternatives: How Do the Prices Compare?

All prices below are the ofox rate on 2026-08-18, per million tokens, so the comparison is single-source and the DeepSeek row is the peak rate rather than the off-peak one.

V4 Flash is 4.5x cheaper than Claude Sonnet 5 on input and 7.6x cheaper on output, and roughly 11x cheaper than Claude Opus 5 or GPT-5.6 Sol on input. It is not the cheapest thing on the list, though: Qwen Max undercuts it on input by about 20%, and only loses on output. Alibaba’s current top-end model is Qwen 3.8 Max, which prices differently again. Anyone telling you DeepSeek is categorically the cheapest option has not checked recently.

The trade-off is the usual one. DeepSeek publishes a concurrency cap rather than an uptime commitment, and instruction-following on nuanced English tasks can lag the frontier models above. For a full capability comparison rather than a price one, see the model comparison guide.

What Limits Should You Know Before You Scale?

A few things that will bite you if you do not know them upfront:

• Concurrency is capped per model, and the two are not close: 2,500 concurrent requests on V4 Flash against 500 on V4 Pro. A Pro-heavy pipeline hits that wall five times sooner.
• Rate limits vary by account tier, and new accounts start conservative. Expect to hit them if you are testing at volume.
• The 1M token context is input-only. Max output is 384K on both models.
• Thinking mode is the default on both models, and the pricing page points at a separate guide for switching it off rather than making it a per-request flag you can guess at.
• FIM completion is beta and works in non-thinking mode only, on both models.
• Prices move, and DeepSeek says so on the pricing page itself: “Product prices may vary and DeepSeek reserves the right to adjust them. We recommend topping up based on your actual usage and regularly checking this page for the most recent pricing information.” After a change of the size that landed on 2026-08-16, that is not boilerplate.

The Bottom Line

At $0.22 per million input tokens off-peak, with a 97% discount on cached prefixes, DeepSeek V4 Flash remains one of the cheapest capable models you can call. The thing that changed in August is that the bill now depends on the clock: the same job costs twice as much between 01:00 and 04:00 or 06:00 and 10:00 UTC. Structure your prompts for cache hits, move batch work out of those seven hours, and reserve V4 Pro for the work that actually needs it.

For a detailed real-world cost analysis using DeepSeek V4 as a Claude Code backend, including cache hit realism and quality trade-offs, see our DeepSeek V4 in Claude Code cost breakdown. For what V4 Pro 0813 actually delivers for its three-times price, see the V4 Pro 0813 breakdown. Get started at platform.deepseek.com, or use ofox.ai if you want a single key that covers DeepSeek alongside Claude, GPT and Gemini.

Six costs the rate card does not show covers the mechanisms behind these rates, peak-hour billing included.

References

• DeepSeek models and pricing
• DeepSeek first API call
• DeepSeek rate limits and isolation
• DeepSeek context caching
• DeepSeek thinking mode
• ofox model page: DeepSeek V4 Flash
• ofox model page: DeepSeek V4 Pro