DeepSeek API Pricing (November 2026)

DeepSeek bills every token at one of two rates, set by the hour the request runs. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday. Every other hour is off-peak at half the peak rate, and so are weekends and Chinese public holidays in full.

USD per 1M tokens. The price table at the top of this page stores the peak rate, because DeepSeek describes off-peak as half of the peak rates.

Peak time is China's working day: 09:00 to 12:00 and 14:00 to 18:00 in Beijing. In New York peak ends at 5:00 a.m. in winter and 6:00 a.m. in summer, so the US working day bills off-peak. The first window does land on US evenings, from 8:00 to 11:00 p.m. Eastern in winter, Sunday to Thursday. In Central Europe the second window covers 7:00 to 11:00 a.m. in winter.

That leaves 35 peak hours in a 168-hour week. DeepSeek's pricing page lists no Batch API discount, so the off-peak rate is how you pay less for evals, backfills and other work that can wait. The off-peak pricing entry explains how other vendors use time-of-day rates.

Context caching is on for every account and needs no code change. When a request repeats a prefix that DeepSeek has already stored on disk, those input tokens bill at the cache-hit rate. The rest bill as a cache miss, at the full input rate.

• No write or storage fee. DeepSeek's pricing page lists only the hit and miss rates. Anthropic and OpenAI charge for writing into the cache on their newer models.
• Whole prefix units only. A hit needs a full match with a stored prefix unit. DeepSeek stores one at the end of each request's input and output, at common prefixes it detects across requests, and at fixed intervals inside long inputs.
• Best effort. DeepSeek does not guarantee a hit rate. A cache nobody uses clears on its own, usually within a few hours to a few days.
• Visible per request. Each response reports prompt_cache_hit_tokens and prompt_cache_miss_tokens in its usage block.
• Per-user isolation. Passing a user_id keeps the cache of each of your end users separate.

The cache matters most for agents and chat, which resend the same system prompt and history on every turn. On V4.1 Flash a hit costs $0.006 per million at peak, against $0.30 for a miss. The prompt caching entry compares how the major APIs price it.

The DeepSeek API runs on a prepaid balance. You top up on the platform, and each request is deducted from that balance. DeepSeek's FAQ says it has one pricing standard and no tiered plans.

• Two balances. An account can hold a topped-up balance and a granted balance. Charges come out of the granted balance first.
• No expiry on top-ups. Topped-up funds do not expire, and an unused balance can be refunded from the Billing page.
• A hard stop at zero. When the balance runs out, the API returns 402 Insufficient Balance until you add funds. Nothing bills beyond the balance.

The bill for a request is three products added together: cache-miss input tokens at the miss rate, cache-hit tokens at the hit rate, and output tokens at the output rate. Reasoning tokens from thinking mode are counted inside the output tokens. Images on V4.1 Flash count as input, at up to 1,024 tokens per image. The LLM API cost calculator runs this sum for a monthly volume.

DeepSeek's published model limit is concurrency, not tokens per minute. Each account can hold 2,500 requests in flight on deepseek-flash and 500 on deepseek-v4-pro. The count covers every API key on the account, and a request holds its slot until the response is complete.

Requests above the limit get HTTP 429. DeepSeek raises the limit on request, matched to your actual traffic, and says the increase costs nothing extra. After an increase, the same per-model limit also applies to each user_id you pass. Search allows 100 requests per second and Fetch 25, counted apart from model traffic. See rate limits for how other APIs meter traffic.

Since October 17, 2026, DeepSeek sells two tools next to its models. The Search API returns web results from DeepSeek's own search engine. The Fetch API returns a single web page as Markdown. Each costs $0.75 per 1,000 successful requests, billed per request rather than per token.

• A request bills only when it returns HTTP 200 with a result. Failed requests, empty results and timeouts are free.
• The count parameter sets how many results come back and does not change the price.
• Since 16:00 UTC on October 31, 2026, calls made by server-side tools in DeepSeek Harness, Claude Code and Codex bill the same way.

The model name in each request decides what answers it, and DeepSeek changed those names twice in 2026. The table shows what each ID does today.

Checkpoint names as printed in the MODEL VERSION row of DeepSeek's pricing table.

DeepSeek calls the routing of the two retired V4 Flash IDs temporary, so code that still sends them should move to deepseek-flash. Until April 24, 2026, deepseek-chat and deepseek-reasoner served the V3 line, V3.2 at the end, in non-thinking and thinking mode. The retired models table keeps their last prices.