DeepSeek API Pricing (2026): Peak/Off-Peak Rates, Payment & Setup

Last updated: 2026-08-31. All prices below are captured as of the dates noted — vendor pricing changes often, always verify on the official pricing page before buying.

DeepSeek's models (V4 Flash, V4 Pro) are among the most cost-effective LLMs you can call in 2026. But two things trip up international developers: DeepSeek's peak/off-peak billing, and the China-only signup friction (Chinese phone number, Chinese payment methods). This guide explains both, shows you what the numbers actually look like, and gives you a working setup that avoids the phone-number problem entirely.

1. How DeepSeek's official pricing works (peak vs off-peak)

DeepSeek bills per 1M tokens with two rate tiers:

• Peak hours: 01:00–04:00 and 06:00–10:00 UTC, Monday–Friday
• Off-peak hours: everything else (including all of Saturday and Sunday)

Off-peak rates are half of peak rates. That means your bill depends on when you call, not just what you call.

Official rates (captured 2026-09-10 from https://api-docs.deepseek.com/quick_start/pricing; preferred model deepseek-flash → DeepSeek-V4.1-Flash; Pro still listed as DeepSeek-V4-Pro-0813):

Model	Input (cache miss) off-peak	Input (cache miss) peak	Output off-peak	Output peak
deepseek-flash (V4.1)	$0.15 / 1M	$0.30 / 1M	$0.60 / 1M	$1.20 / 1M
deepseek-v4-pro *	$0.66 / 1M	$1.32 / 1M	$1.98 / 1M	$3.96 / 1M

Cache hits are dramatically cheaper (Flash input with cache hit: $0.003 off-peak / $0.006 peak per 1M). Both list 1M context, up to 384K max output, tool calls, JSON, and the Responses API. Flash also supports vision.

Two honest notes:

1. V4.1 Flash is cheaper than the mid-August V4 Flash sheet ($0.22/$0.66 → $0.15/$0.60 off-peak). Legacy deepseek-v4-flash still works but is served as V4.1 Flash.
2. From 12:00 Beijing time on 2026-09-14, official docs say deepseek-v4-pro requests are routed to V4.1 Flash and billed at Flash rates until V4.1 Pro ships. Peak windows remain UTC Mon–Fri 01:00–04:00 & 06:00–10:00.

2. The problem for international users: China-only signup

DeepSeek's first-party platform requires:

• A Chinese mobile number for registration
• Chinese payment methods (Alipay/WeChat or China-issued cards) for top-up

If you're outside China — or simply don't have a Chinese phone number — you can't directly use the first-party platform. That's the exact gap third-party access exists for: AI gateways / aggregators that resell or route DeepSeek capacity and accept PayPal or international cards.

3. The three-column price comparison

"Which access is cheapest?" depends on when you call and what you run. Here's an honest three-way snapshot (all captured 2026-08-31):

Access	Input / 1M	Output / 1M	Notes
DeepSeek official — Flash off-peak	$0.22 (miss)	$0.66	Requires Chinese phone + CNY payment. Peak = 2× (see §1)
DeepSeek official — Flash peak	$0.44 (miss)	$1.32	Peak hours: 01:00–04:00 / 06:00–10:00 UTC, Mon–Fri
DeepSeek official — Pro off-peak	$0.66 (miss)	$1.98	Higher concurrency, stronger reasoning
DeepSeek official — Pro peak	$1.32 (miss)	$3.96	—
OpenRouter — deepseek/deepseek-v4-pro	$1.04	$2.07	Blended flat rate; source: OpenRouter model API, 2026-08-31
OpenRouter — deepseek/deepseek-v4-flash	$0.08	$0.17	Third-party hosted (varies by provider)
DeepInfra — DeepSeek-V4-Flash-0731	$0.08	$0.18	Cached input $0.016; 1M context
DeepInfra — DeepSeek-V4-Pro	$1.30	$2.60	Cached input $0.10
AI24X — flash tier (one key, flat)	from $0.35/M output	—	One flat price, smart routing, no peak/off-peak tracking; prices captured 2026-08-05, see https://www.ai24x.com/pricing.html

How to read this table honestly:

• If you have a Chinese phone and want to micro-manage call times, the official off-peak rate is the cheapest floor — but you pay peak rates when you call during UTC business hours, which is exactly when most developers work.
• OpenRouter / DeepInfra offer flat-rate DeepSeek access with international cards — their flash-tier prices are close to official off-peak, their pro-tier prices sit above official off-peak.
• AI24X's value isn't "cheapest per token ever" — it's that one key covers DeepSeek plus Qwen, GLM, Kimi and international flagships (GPT-5/Claude/Gemini), with smart routing picking the best channel so you don't hand-track five price lists. Flash starts from $0.35/M output.

Prices verified 2026-08-31 against: official DeepSeek docs (https://api-docs.deepseek.com/quick_start/pricing), OpenRouter model API (https://openrouter.ai), DeepInfra (https://deepinfra.com), AI24X pricing (https://www.ai24x.com/pricing.html). If a number matters to your budget, click through and check the live page — prices move.

4. A worked cost example (why "when" and "where" matter)

Let's make the peak/off-peak math concrete. Say a developer sends 100M input tokens (cache miss) and 20M output tokens per month to a Flash-class model:

Scenario	Input bill	Output bill	Total
Official, all peak	100 × $0.44 = $44.00	20 × $1.32 = $26.40	$70.40
Official, all off-peak	100 × $0.22 = $22.00	20 × $0.66 = $13.20	$35.20
Aggregator flash tier (flat)	varies by routing	~20 × $0.35 = $7.00	often the effective bill lands between the two, minus the ops cost of scheduling

The point isn't that any one row always wins — it's that your real cost depends on your call pattern, and most teams don't want to schedule their production traffic around UTC peak windows. A gateway's job is to absorb that complexity: smart routing keeps routine calls on cheap channels automatically, and you get one predictable bill instead of a spreadsheet.

5. One key for DeepSeek + Qwen + GLM + Kimi + flagships

The reason developers use an aggregator instead of juggling five vendor accounts:

• One OpenAI-compatible endpoint: POST /v1/chat/completions with Authorization: Bearer <key>
• One payment rail: PayPal or card — no Chinese phone number required
• All the Chinese model families on one key: DeepSeek, Qwen (Alibaba), GLM (Zhipu), Kimi (Moonshot) — plus GPT-5/Claude/Gemini when a task needs flagship power
• Smart routing: for each request, the gateway evaluates available upstream channels and picks the best one by price, speed, and stability, with automatic failover when a channel degrades. Ask for a specific model when you want it; let routing handle the rest.

That's the "One API. Every AI. Pay Less." model — you stop being a pricing analyst and just call models.

Which model should you actually use?

A practical default ladder for DeepSeek-class workloads in 2026:

Workload	Default	Why
Everyday chat, summarization, light coding	flash	Best reasoning per dollar; cheap enough to use liberally
Serious coding agents, math, long reasoning chains	pro	Stronger reasoning; costs ~3× flash output — escalate only when the task demands it
Long documents, big codebases, agent memory	vip-kimi / 1M-context models	Context length is the deciding factor
Vision/audio, multimodal pipelines	Qwen multimodal / vision models	Breadth of modalities out of the box
Hardest reasoning, high-stakes writing	GPT-5.x / Claude / Gemini	Escalate to flagships only when needed

Keep the default cheap and escalate deliberately — that single habit cuts API bills more than any rate negotiation.

6. Free tier (what's actually available)

For trying things out without paying first: AI24X provides a free shared channel of 100K tokens/day when your balance is empty. That's the current live offering — there is no "register and get 5,000 tokens" promotion, and no signup bonus. Check the console for the exact current policy.

Common pitfalls to avoid

• Old pricing numbers in blog posts. Mid-August 2026 pricing changes mean older "$0.14/M output"-style figures are stale. Always cross-check against the official pricing page on the day you commit.
• Scheduling everything at peak. If you're on the official platform, off-peak (UTC evenings/weekends) literally halves input and output rates for cache-miss traffic. Batch jobs belong in off-peak.
• Ignoring cache hits. DeepSeek's cache-hit input pricing is orders of magnitude cheaper. Prefix caching rewards stable system prompts — keep your prompts stable where possible.
• Assuming "official" means "only option". Third-party gateways are legitimate access paths for international users, but check the transparency of each: does the pricing page show capture dates and sources? Opaque pricing is a red flag.

7. Copy-paste setup

Base URL (OpenAI-compatible)

https://api.ai24x.com/v1

Quick test with curl

curl -X POST "https://api.ai24x.com/v1/chat/completions" \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer sk-..." \
  -d '{"model":"flash","messages":[{"role":"user","content":"Hello! Explain DeepSeek off-peak pricing in one sentence."}]}'

Codex CLI

# ~/.codex/config.toml
model_provider = "ai24x"

[model_providers.ai24x]
name = "AI24X"
base_url = "https://api.ai24x.com/v1"
env_key = "AI24X_API_KEY"
wire_api = "chat"

OpenClaw / OpenClaw-style agent config

{
  "model": "flash",
  "modelProvider": "ai24x",
  "providers": {
    "ai24x": {
      "baseUrl": "https://api.ai24x.com/v1",
      "apiKey": "sk-...",
      "models": ["flash", "pro", "ultra", "vip-qwen", "vip-glm", "vip-kimi"]
    }
  }
}

Then set AI24X_API_KEY=sk-... in your environment. Existing OpenAI SDKs and tools (Open WebUI, LobeChat, Cursor, Dify, LiteLLM, n8n) connect the same way: change base URL + key, keep the rest of your code.

8. FAQ

Are off-peak rates really half of peak?

On DeepSeek's first-party platform, yes — off-peak is 50% of peak (peak = 01:00–04:00 and 06:00–10:00 UTC, Mon–Fri; weekends are off-peak). Verified on the official pricing page 2026-08-31. Aggregators like AI24X don't split billing by hour — one flat rate, routing decides.

Can I pay for DeepSeek access with PayPal or an international card?

Not on DeepSeek's first-party platform (it needs a Chinese phone number and CNY payment). Via third-party gateways — yes: AI24X accepts PayPal and card, no Chinese phone number needed. See the price table in §3 for alternatives.

Is AI24X the official DeepSeek API?

No. AI24X is an independent AI gateway / aggregator (not affiliated with DeepSeek) that routes to DeepSeek, Qwen, GLM, Kimi and international models through one OpenAI-compatible key.

What free quota do I get?

A shared free channel of 100K tokens/day when your balance is empty (subject to current console policy; no signup bonus, no "5,000 free tokens" promotion).

What's the refund policy?

Refund rules are set per platform. AI24X's billing is prepaid-credit based; check the console's billing/refund terms for the current policy, and top up only what you plan to use.

9. Getting started

1. Register at https://www.ai24x.com and create an API key in the console.
2. Top up at console → billing (PayPal or card).
3. Use the base_url + key from the Copy-paste setup section in any OpenAI-compatible tool — done.

Developer docs and integrations: https://www.ai24x.com/docs.html · Console: https://www.ai24x.com/console.html · Pricing: https://www.ai24x.com/pricing.html

Educational/informational article. This page is not affiliated with DeepSeek and does not constitute investment advice of any kind. All prices are as captured on the dates noted and may change; always check official pricing pages.

Related: OpenAI-compatible API · One API for multiple AI models · BYOK AI Gateway · AI API cost basics