DeepSeek — $0.15 per million input tokens and $0.60 per million output off-peak on deepseek-flash — with off-peak rates set at exactly half of peak.
Short answer: DeepSeek bills per token with no plans and no monthly fee. On deepseek-flash, one million cache-miss input tokens cost $0.15 off-peak and $0.30 at peak, cached input costs $0.003 off-peak and $0.006 at peak, and one million output tokens cost $0.60 off-peak and $1.20 at peak. On deepseek-v4-pro the same four figures are $0.66 and $1.32 for cache-miss input, $0.022 and $0.044 for cache hits, and $1.98 and $3.96 for output. Off-peak rates are exactly half of peak, and peak hours are 01:00–04:00 and 06:00–10:00 UTC Monday to Friday, excluding Chinese public holidays. Both models carry a 1M-token context window and a 384K maximum output. Concurrency is 2,500 on Flash and 500 on Pro. Charges come off a topped-up or granted balance. Read from api-docs.deepseek.com on 1 October 2026.
On this page
Verdict at a glance
Prompt caching at $0.003 per million input tokens off-peak is fifty times cheaper than a cache miss, which makes a long system prompt or a fixed document set effectively free to re-send — and that is where most agent workloads spend their input budget.
The headline price depends on the clock. Peak hours are 01:00–04:00 and 06:00–10:00 UTC on weekdays, and anything running then costs exactly double the off-peak rate, so a European or US business-hours workload may sit permanently in the expensive window. There are no plans, no SLA tier and no committed-capacity option — just a prepaid balance and a note that DeepSeek reserves the right to adjust prices.
DeepSeek sells two models and one billing mechanism. deepseek-flash, running DeepSeek-V4.1-Flash, costs $0.15 per million cache-miss input tokens and $0.60 per million output tokens off-peak, doubling to $0.30 and $1.20 at peak. deepseek-v4-pro, running DeepSeek-V4-Pro-0813, costs $0.66 and $1.98 off-peak and $1.32 and $3.96 at peak. Cached input is the line worth staring at: $0.003 per million tokens off-peak on Flash and $0.022 on Pro, a fiftieth and a thirtieth of the respective cache-miss rates. For an agent that re-sends the same instructions and the same reference material on every call, that is the difference between a token bill that scales with conversation length and one that barely moves. Both models offer a 1M-token context window and up to 384K output tokens, JSON output, tool calls, a Responses API and an Anthropic-format endpoint at api.deepseek.com/anthropic — which matters, because DeepSeek documents integrations with Claude Code, Codex, OpenCode and its own DeepSeek Harness, so existing agent tooling points at it without a rewrite. Flash supports vision and thinking plus non-thinking modes; Pro supports thinking by default but not vision, and FIM completion is non-thinking only on both. Concurrency limits are 2,500 for Flash and 500 for Pro. The timing rule is the real catch. Peak hours are 01:00–04:00 and 06:00–10:00 UTC Monday to Friday, excluding Chinese public holidays in full, and everything else — including weekends — is off-peak at half the price. A workload scheduled around that saves 50%; a workload that cannot be rescheduled pays double the number in every comparison table. Billing is prepaid against a topped-up or granted balance, granted balance consumed first, with no plans, no SLA tier and an explicit note that prices may be adjusted. Four stars: the rates are among the lowest published for frontier-class context lengths, caching is aggressively cheap, and the Anthropic-format endpoint removes most migration friction. The deductions are for peak pricing that penalises Western business hours and for the absence of any committed-capacity or support tier.
Quick facts
| Attribute | Detail |
|---|---|
| Ideal team | Developers running high-volume or long-context LLM workloads |
| Learning curve | Low — OpenAI and Anthropic-format endpoints |
| Setup time | Minutes |
| Free trial | ⚠️ Granted balance where offered; no stated free tier |
| Free tier | ⚠️ None published |
| Entry | $0.15 per 1M input tokens off-peak (deepseek-flash, cache miss) |
| Context window | 1M tokens, 384K maximum output |
| Peak pricing | ⚠️ Double off-peak, 01:00–04:00 and 06:00–10:00 UTC weekdays |
Overview
There are two models and no plans. deepseek-flash is DeepSeek-V4.1-Flash: $0.003 per million cached input tokens off-peak and $0.006 at peak, $0.15 and $0.30 per million cache-miss input tokens, $0.60 and $1.20 per million output tokens, with a concurrency limit of 2,500. deepseek-v4-pro is DeepSeek-V4-Pro-0813: $0.022 and $0.044 cached, $0.66 and $1.32 cache-miss input, $1.98 and $3.96 output, concurrency 500. The legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted but those models are retired; requests are served by V4.1-Flash and billed at the Flash price.
Capability is close to identical across the two. Both expose a 1M-token context window with a maximum output of 384K tokens, JSON output, tool calls, a Responses API, an Anthropic-format API and beta chat-prefix completion. Flash supports both non-thinking and thinking modes with thinking as the default, and supports vision; Pro supports thinking by default but not vision. FIM completion is available in non-thinking mode only on both. Two base URLs are published: api.deepseek.com in OpenAI format and api.deepseek.com/anthropic in Anthropic format, and the docs list agent integrations including DeepSeek Harness, Claude Code, Codex, OpenCode, OpenClaw, Hermes, Reasonix, WorkBuddy/CodeBuddy and Qoder.
The pricing mechanic that changes the bill is time of day. The docs state that off-peak rates are half of peak rates, that peak hours are 01:00–04:00 and 06:00–10:00 UTC Monday through Friday excluding Chinese public holidays, and that all other hours — including weekends and Chinese public holidays in full — are off-peak. Billing is straightforward: expense equals tokens multiplied by price, deducted from a topped-up or granted balance with granted balance used first. The page advises topping up to match actual usage and checking back for current pricing, and states that DeepSeek reserves the right to adjust prices. There is no published free tier, no plan ladder, no SLA tier and no committed-capacity discount.
Key features
- deepseek-flash$0.15/1M input, $0.60/1M output off-peak
- deepseek-v4-pro$0.66/1M input, $1.98/1M output off-peak
- Prompt caching$0.003/1M cached input tokens off-peak on Flash
- 1M-token contexton both models, 384K maximum output
- Anthropic-format endpointapi.deepseek.com/anthropic
- Tool calls and JSON outputon both models
- VisionFlash only; Pro does not support it
- Thinking modesswitchable on Flash, default on Pro
- Concurrency2,500 on Flash, 500 on Pro
- Peak and off-peak rates⚠️ peak is exactly double
Pricing (as of 2026-10-01)
Model: per million tokens, prepaid against a balance; off-peak rates are half of peak; no plans and no monthly fee · Free tier: ⚠️ None published — a granted balance is consumed before topped-up credit where one exists · Entry paid: $0.15 per 1M input tokens off-peak (deepseek-flash, cache miss)
| Plan | Price | Includes |
|---|---|---|
| deepseek-flash — input (cache miss) | $0.15/1M off-peak | $0.30/1M at peak · DeepSeek-V4.1-Flash · 1M context · vision supported |
| deepseek-flash — input (cache hit) | $0.003/1M off-peak | $0.006/1M at peak — a fiftieth of the cache-miss rate |
| deepseek-flash — output | $0.60/1M off-peak | $1.20/1M at peak · 384K maximum output · concurrency 2,500 |
| deepseek-v4-pro — input (cache miss) | $0.66/1M off-peak | $1.32/1M at peak · cache hit $0.022 off-peak and $0.044 at peak |
| deepseek-v4-pro — output | $1.98/1M off-peak | $3.96/1M at peak · concurrency 500 · no vision support |
⚠️ Off-peak rates are exactly half of peak. Peak hours are 01:00–04:00 and 06:00–10:00 UTC Monday to Friday, excluding Chinese public holidays; everything else including weekends is off-peak.
Prompt caching is the largest lever on the bill: $0.003 per million cached input tokens off-peak on Flash against $0.15 for a cache miss, and $0.022 against $0.66 on Pro.
Billing is prepaid. Expense equals tokens multiplied by price, deducted from a topped-up or granted balance, with granted balance used first. The docs advise topping up to match actual usage.
The legacy model names deepseek-v4-flash and deepseek-v4-flash-vision-exp still resolve, but those models are retired and requests are served by DeepSeek-V4.1-Flash at the Flash price.
⚠️ There is no plan ladder, no SLA tier and no committed-capacity discount, and the docs state that DeepSeek reserves the right to adjust prices — so model this as a variable cost, not a contract.
Ratings breakdown
Pros
- Among the lowest published per-token rates for a 1M-token context
- Cached input at $0.003 per million tokens off-peak on Flash
- Anthropic-format endpoint means existing agent tooling points at it unchanged
- Documented integrations with Claude Code, Codex and OpenCode
- Concurrency limit of 2,500 on the Flash model
- No monthly fee and no plan commitment
Cons
- ⚠️ Peak pricing doubles the rate during European and Asian business hours
- ⚠️ No SLA tier or committed capacity
- No published free tier
- Pro model does not support vision
- FIM completion limited to non-thinking mode
- Docs explicitly reserve the right to change prices
Who it’s for
- Teams running batch or asynchronous workloads that can be scheduled off-peak
- Agent builders who re-send large fixed prompts and benefit from cheap caching
- Developers wanting a 1M-token context at a low per-token rate
- Projects already written against the OpenAI or Anthropic SDK
Who it’s not for
- Workloads that must run in the 01:00–04:00 or 06:00–10:00 UTC windows at lowest cost
- Buyers who need an SLA, support contract or committed capacity
- Applications requiring vision on the strongest available model
- Organisations with data-residency requirements they have not checked
Best use cases
- Running an overnight document-processing batch at off-peak rates
- Serving an agent with a 100K-token system prompt entirely from cache
- Drop-in substitution behind code written for the Anthropic SDK
- High-concurrency classification at 2,500 parallel requests
- Long-context summarisation across a 1M-token corpus
Integrations
The docs list agent integrations for DeepSeek Harness, Claude Code, Codex, OpenCode, OpenClaw, Hermes, Reasonix, WorkBuddy/CodeBuddy and Qoder, and publish both an OpenAI-format and an Anthropic-format base URL. That makes it a substitution target rather than a new integration for anything already built against Claude or OpenAI, and it slots into orchestration built in LangChain or n8n.
Alternatives
OpenAI and Claude are the direct comparisons on capability and both publish higher per-token rates with support and SLA options DeepSeek does not offer. Gemini competes on long context and bundles access into consumer subscriptions. For code work specifically, compare against Claude Code, Codex and Cursor, which sell seats rather than tokens.
Recent changes
Read from api-docs.deepseek.com on 1 October 2026. The catalogue is now two models: deepseek-flash, running DeepSeek-V4.1-Flash, and deepseek-v4-pro, running DeepSeek-V4-Pro-0813. The legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp still resolve but those models are retired and their requests are served by V4.1-Flash at the Flash price. Current rates per million tokens: Flash $0.003 cached and $0.15 cache-miss input and $0.60 output off-peak, doubling at peak; Pro $0.022 cached and $0.66 cache-miss input and $1.98 output off-peak, doubling at peak. Both models now carry a 1M-token context window with a 384K maximum output, tool calls, JSON output, a Responses API and an Anthropic-format endpoint; Flash supports vision and Pro does not. Peak hours are 01:00–04:00 and 06:00–10:00 UTC Monday to Friday excluding Chinese public holidays, with all other hours off-peak. Concurrency limits are 2,500 for Flash and 500 for Pro.
Frequently asked questions
How much does the DeepSeek API cost?
On deepseek-flash, $0.15 per million cache-miss input tokens and $0.60 per million output tokens off-peak, doubling to $0.30 and $1.20 at peak. On deepseek-v4-pro, $0.66 and $1.98 off-peak and $1.32 and $3.96 at peak. Read on 1 October 2026.
What are DeepSeek’s peak hours?
01:00–04:00 and 06:00–10:00 UTC, Monday through Friday, excluding Chinese public holidays. Everything else — including weekends and Chinese public holidays in full — is off-peak at half the price.
How much does prompt caching save?
A great deal. Cached input is $0.003 per million tokens off-peak on deepseek-flash against $0.15 for a cache miss, and $0.022 against $0.66 on deepseek-v4-pro — a fiftieth and a thirtieth of the respective rates.
Does DeepSeek have a free tier?
No published free tier. Billing runs off a prepaid balance, and where a granted balance exists it is consumed before topped-up credit. The docs advise topping up based on actual usage.
What is DeepSeek’s context window?
One million tokens on both deepseek-flash and deepseek-v4-pro, with a maximum output of 384K tokens. Concurrency limits are 2,500 requests on Flash and 500 on Pro.
Can I use DeepSeek with Anthropic-format code?
Yes. The docs publish api.deepseek.com/anthropic as an Anthropic-format base URL alongside the OpenAI-format api.deepseek.com, and list agent integrations including Claude Code, Codex and OpenCode.
Sources
- DeepSeek API docs – models and pricing table (read 2026-10-01)
- DeepSeek API docs – peak and off-peak hours and deduction rules (read 2026-10-01)
- DeepSeek API docs – rate limits and concurrency isolation (read 2026-10-01)
Compiled from public sources; last verified 2026-10-01. Pricing is subject to change — verify with the vendor. Ratings reflect editorial opinion based on documented capabilities, published pricing, and aggregated user reviews. DigiJaws does not accept payment for review placement or rating.
Looking for the numbers? Our DeepSeek pricing page breaks down every plan, what each upgrade costs over a year, and what the headline price leaves out.
DeepSeek Review (4/5): API Pricing (2026). DigiJaws, 2026. https://digijaws.com/reviews/deepseek-review/