DeepSeek API Pricing (Oct 2026): Peak & Off-Peak Token Rates

1. Official DeepSeek API Pricing Table (Checked 2026-10-27)

Verified on 2026-10-27 against DeepSeek's official Models & Pricing and Chinese Pricing documentation (quoted per 1,000,000 tokens / 1M tokens), active billing centers on deepseek-flash (DeepSeek-V4.1-Flash-1017) and deepseek-v4-pro (DeepSeek-V4-Pro-0813), while legacy IDs deepseek-chat and deepseek-reasoner were retired on the official DeepSeek service after 2026-07-24 15:59 UTC. For deepseek-flash, Off-Peak rates are $0.003 / 1M input cache hit (0.02 CNY), $0.15 / 1M input cache miss (1 CNY), and $0.60 / 1M output (4 CNY), doubling during Peak hours to $0.006 / 1M (0.04 CNY), $0.30 / 1M (2 CNY), and $1.20 / 1M (8 CNY):

Official Built-in Tools & Tariff Rules

• Built-in Tool APIs (Search and Fetch): Priced at $0.75 (5 CNY) per 1,000 successful requests on the official platform. During the initial launch period through DeepSeek Harness, Claude Code, and Codex, Search and Fetch calls are free until 00:00 Beijing time on November 1, 2026, after which per-request billing applies.
• USD and CNY are separately published official reference denominations: DeepSeek publishes distinct USD (en) and CNY (zh-cn) price schedules rather than applying a single uniform FX conversion factor across every row.
• Shared 1M context window and 384K max output limit: Both deepseek-flash (DeepSeek-V4.1-Flash-1017) and deepseek-v4-pro (DeepSeek-V4-Pro-0813) support a 1M context window, up to 384K max output tokens, non-thinking and thinking modes (thinking enabled by default), JSON Output, Tool Calls, Chat Prefix Completion (Beta), and FIM Completion (Beta, non-thinking mode only).

---

2. UniRoute vs. Official DeepSeek Price Comparison (Peak vs. Off-Peak)

USD per 1M tokens; official checked: 2026-10-27. UniRoute retail: live gateway snapshot. Saving = 1 − retail / official, rounded to a whole percent; + means more expensive, 0% means equal at displayed precision.Standard and Discounted describe price bands, not upstream providers. Where routes share a band, one complete lowest default public rate set is shown; your key may use another route or account rate. A snapshot is not a frozen billing contract. Comparison covers ordinary standard context, excluding long-context surcharges and priority processing.

Model	Official input	Official output	UniRoute input	UniRoute output	You save (input / output)
deepseek-flash · Discounted price band · Peak	$0.30	$1.20	$0.225	$0.90	25% / 25%
deepseek-flash · Discounted price band · Off-peak	$0.15	$0.60	$0.225	$0.90	+50% / +50%
deepseek-flash · Standard price band · Peak	$0.30	$1.20	$0.30	$1.20	0% / 0%
deepseek-flash · Standard price band · Off-peak	$0.15	$0.60	$0.30	$1.20	+100% / +100%
deepseek-v4-pro · Discounted price band · Peak	$1.32	$3.96	$0.99	$2.97	25% / 25%
deepseek-v4-pro · Discounted price band · Off-peak	$0.66	$1.98	$0.99	$2.97	+50% / +50%
deepseek-v4-pro · Standard price band · Peak	$1.32	$3.96	$1.32	$3.96	0% / 0%
deepseek-v4-pro · Standard price band · Off-peak	$0.66	$1.98	$1.32	$3.96	+100% / +100%

Peak hours are weekdays 09:00–12:00 and 14:00–18:00 Beijing time (01:00–04:00 and 06:00–10:00 UTC), excluding Chinese public holidays; all other times are off-peak. On DeepSeek's official API, the legacy name deepseek-v4-flash is billed at the Flash price.

How to Choose Between Official Direct API and UniRoute (Honest Breakdown)

DeepSeek's official direct platform bills by time of day, whereas UniRoute bills a flat 24/7 rate (see Channel Management for tier selection and failover):

1. Official Peak Hours (01:00–04:00 UTC and 06:00–10:00 UTC / Beijing 09:00–12:00 and 14:00–18:00, Mon–Fri excluding Chinese public holidays):
   • Official deepseek-flash Peak rate is $0.30 input / $1.20 output per 1M tokens.
   • As of 2026-10-27, UniRoute's Discounted Tier (Discounted) is $0.225 input / $0.90 output (25% cheaper than official Peak), and UniRoute's Standard Tier (Standard) is $0.30 input / $1.20 output (same as the official Peak rate, with automatic failover to the Standard Tier when discounted channels are unavailable)—always refer to the live comparison table above for current real-time rates.
2. Official Off-Peak Hours (all other hours, weekday lunch 12:00–14:00 Beijing time, nights, and all day Saturday/Sunday & Chinese public holidays):
   • Official deepseek-flash Off-Peak rate drops by 50% to $0.15 input / $0.60 output per 1M tokens, which is lower than UniRoute's 24/7 flat rate ($0.225 / $0.90 on Discounted as of 2026-10-27).
3. Which option fits your workload?
   • Choose UniRoute if your team calls DeepSeek primarily during daytime working hours (Beijing 09:00–12:00 and 14:00–18:00): You save 25% against official Peak on the Discounted Tier, avoid writing time-window scheduling logic, and use one unified API key and one USD balance to call DeepSeek alongside Claude (Discounted Tier as low as 42% of official) and OpenAI GPT (Discounted Tier as low as 28% of official), billed in USD from your UniRoute balance.
   • Choose Official DeepSeek Direct API if your workload is mostly overnight or weekend batch processing: Running asynchronous batch jobs directly against DeepSeek during Off-Peak hours ($0.15 / $0.60) is more cost-effective.

---

3. Interactive DeepSeek API Cost Calculator

Select the official Peak or Off-Peak billing window and a model, then enter input tokens per request, output tokens per request, and requests per day to compare official DeepSeek rates against UniRoute:

---

4. Peak vs. Off-Peak Pricing Windows (Official Direct API Only)

DeepSeek divides the week into Peak and Off-Peak windows based on UTC (UTC+8 Beijing time):

---

5. Model IDs, Legacy Aliases, and the deepseek-v4-pro Status Note

DeepSeek's model naming changed significantly across the V3.2 and V4/V4.1 transitions. Using stale model strings from older tutorials or SDK defaults can either fail on the official direct API or route to a different catalog SKU than you intended.

---

6. How DeepSeek Token Billing Works: Cache Hits, Cache Misses, and Thinking Tokens

1. The Exact Formula (No Double-Counting)

DeepSeek's response usage object splits input tokens into two mutually exclusive buckets (prompt_cache_hit_tokens and prompt_cache_miss_tokens) and includes reasoning tokens inside completion_tokens:

prompt_tokens = prompt_cache_hit_tokens + prompt_cache_miss_tokens

request_cost =
  (prompt_cache_hit_tokens / 1,000,000) * input_cache_hit_rate +
  (prompt_cache_miss_tokens / 1,000,000) * input_cache_miss_rate +
  (completion_tokens / 1,000,000) * output_rate

2. Mapping usage Fields Correctly

3. Context Caching on Disk: Automatic, Best-Effort, and Prefix-Based

According to DeepSeek's Context Caching documentation, context caching is enabled by default for all users on the official direct service without code changes or extra request headers:

• Complete persisted prefix-unit reuse required: Cache hits occur only when the beginning (prefix) of a subsequent prompt completely reuses a persisted prefix storage unit on disk—sharing a short leading substring is not enough if it does not span a full persisted unit.
• Best-effort construction latency: Building a cache entry on disk takes a few seconds after a request completes, and only persisted storage units can be matched.
• Automatic expiration: Unused cache entries are cleared automatically over time, and DeepSeek does not guarantee a 100% cache hit rate.
• Published rate categories vs. missing gateway fields: DeepSeek's official pricing reference publishes three rate categories—Input (Cache Hit), Input (Cache Miss), and Output—and does not list a separate cache-write rate row. When inspecting third-party gateway catalogs, if fields like cache_write_1h_price_per_million_usd or cache_write_5m_price_per_million_usd are null or omitted, treat those missing categories as unpublished/unknown.

4. Thinking Mode Parameters on the Official Direct API

As documented in DeepSeek's Thinking Mode guide and Create Chat Completion reference, both deepseek-flash and deepseek-v4-pro enable thinking mode by default on the official direct API (thinking: {"type": "enabled"} with top-level reasoning_effort: "high"):

---

7. 3 Reproducible Hypothetical Cost Examples (Official Direct Rates)

Scope and billing assumptions: The three scenarios below are hypothetical arithmetic illustrations modeling a single request to one provider (the official DeepSeek direct API) landing entirely within one time band (either 100% Off-Peak or 100% Peak) under the 2026-10-27 official USD schedule.

---

8. Calling DeepSeek Through UniRoute: Live Verification Workflow

If your team routes DeepSeek alongside OpenAI, Anthropic, and Google models through UniRoute, keep in mind that official DeepSeek direct API keys, account balances, and bills are completely separate from UniRoute API keys, wallet balances, and usage records. Follow this three-step verification workflow:

1. Verify the Exact Model SKU on /models: Start at the public UniRoute Model Catalog (/models) or inspect the DeepSeek V4.1 Flash model page (/models/deepseek-v4-1-flash).
2. Configure Your API Key and Route Selection: Create a scoped key via the Create API Key guide (/docs/create-api-key) or in your API Keys Console (/console/api-keys), and review Route Selection (/docs/route-selection).
3. Send One Small Test Call and Inspect the Usage Console: Run a single test call against https://api.uniroute.app/v1, then open the UniRoute Console (/console) to inspect the request detail drawer and verify the routed channel, token breakdown, and recorded actual cost.

---

9. Frequently Asked Questions (FAQ)

Are deepseek-chat and deepseek-reasoner still available on DeepSeek's official API in 2026?

No. According to DeepSeek's 2026-04-24 official announcement, both deepseek-chat and deepseek-reasoner were retired on the official DeepSeek service after 2026-07-24 15:59 UTC and are no longer accessible there. Direct API callers should migrate to deepseek-flash (DeepSeek-V4.1-Flash-1017) or deepseek-v4-pro.

Do I pay extra for reasoning_tokens on top of completion_tokens?

No. In DeepSeek's API response, completion_tokens is the total output token count and already includes both reasoning_content (completion_tokens_details.reasoning_tokens) and the final visible content. Multiply completion_tokens by the output rate once—never add reasoning_tokens on top of completion_tokens.

Why did my repeated prompt not get a 100% cache hit rate?

DeepSeek's context caching is automatic and prefix-based, and a subsequent request must completely reuse a persisted prefix storage unit on disk to hit the cache. Keep static instructions, schemas, and few-shot examples at the very start of your messages array (A + B + C reusing persisted A + B), and avoid mutating timestamps or user IDs in the system prompt prefix.

---

Related API Pricing & Routing Guides

• OpenAI API Pricing (Oct 2026): Every GPT Model per 1M Tokens
• Claude Code API Pricing: Token Rates & Cost Estimates (2026)
• OpenAI Responses API vs. Chat Completions: Migration, State & Routing Guide