DeepSeek API pricing in 2026: peak, off-peak and cache

How much do DeepSeek models cost per million tokens?

DeepSeek bills input and output tokens, with a much lower rate for input it has cached. Off-peak rates are exactly half of peak rates. Figures come from DeepSeek’s Models and Pricing page. Prices checked November 10, 2026.

Both models have a 1 million token context and up to 384K output tokens. Thinking mode is on by default. Vision works on deepseek-flash only. The legacy name deepseek-v4-flash still works, but V4.1 Flash serves it at the Flash price.

When are DeepSeek’s off-peak hours?

Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday, excluding Chinese public holidays. Every other hour is off-peak, including weekends and Chinese public holidays in full. That makes about 79% of a normal week off-peak.

In US Eastern Standard Time, peak runs 8 to 11 p.m. Sunday through Thursday and 1 to 5 a.m. Monday through Friday. A US team working daytime hours pays off-peak rates for nearly all its traffic. Daylight saving time shifts these by one hour.

What does a real DeepSeek workload cost?

Take 10,000 calls, each with a 2,000-token prompt and a 500-token answer: 20 million input tokens and 5 million output tokens. The cached rows assume 1,500 tokens of each prompt hit the cache, such as a fixed system prompt.

Thinking mode is the default, and the reasoning the model writes counts toward its output tokens. If answers include long reasoning, output tokens and cost can be several times the visible answer. Switch to non-thinking mode for simple tasks.

What else affects a DeepSeek bill?

DeepSeek bills from a prepaid balance. Any granted balance is spent before money you topped up. DeepSeek says prices may change and recommends topping up based on actual usage.

• Search and Fetch APIs: $0.75 per 1,000 successful requests each, billed apart from tokens.
• Concurrency: 2,500 simultaneous requests for deepseek-flash and 500 for deepseek-v4-pro per account. Above that, requests fail with HTTP 429. DeepSeek raises limits on request at no cost.
• Fill-in-the-middle completion works in non-thinking mode only.
• Both OpenAI-format (https://api.deepseek.com) and Anthropic-format (https://api.deepseek.com/anthropic) base URLs are available.

How is DeepSeek billed through Locus Pro?

Locus Pro offers deepseek/chat and deepseek/fim, paid from the same prepaid credit balance as every other provider, with no DeepSeek account or top-up. Both take OpenAI-format requests, with thinking on by default. Catalog checked November 10, 2026.

Each call places a hold sized from the prompt and max_tokens, then settles to the tokens DeepSeek reports. Send max_credits to cap any call, and set a monthly hard limit to stop new calls once you reach it.

• DeepSeek pay per use through Locus Pro
• Browse model endpoints

Which DeepSeek option should you pick?

DeepSeek’s list rates are already low, so the main savings come from timing and caching. Run batch work off-peak, keep long system prompts identical so they hit the cache, and use Flash unless a task needs V4 Pro.

For steady volume, DeepSeek can cost less per call with your own DeepSeek account and API key. Locus Pro fits when an agent routes some steps to DeepSeek and others to Claude, Gemini or paid data APIs, and you want one balance and a price ceiling per call.

• Chat, extraction and summaries at volume: deepseek-flash, off-peak where you can.
• Harder reasoning and code: deepseek-v4-pro, at about 3 to 4 times the Flash price.
• Image inputs: deepseek-flash, since V4 Pro does not support vision.
• Copilot-style code completion: fill-in-the-middle in non-thinking mode.

• Cheapest way to call AI models by API