DeepSeek API review
by DeepSeekLLM APIs
Per-token access to DeepSeek's V4.1-Flash and V4 Pro models in OpenAI and Anthropic request formats, with half-price off-peak hours.
- Paid only
- Self-serve
- Usage-based
- Public prices
What you're paying for
- Worth the price (main value)
- Easy to adopt
A 1M-token-context model billed per token, with automatic cache hits at 98% off and half-price off-peak hours.
DeepSeek charges $0.30 per 1M input and $1.20 per 1M output tokens for deepseek-flash at the peak rate, and half that off-peak. Context caching is on for every account, and cache hits cost $0.006 per 1M tokens, so agents and long multi-turn chats that resend the same prefix pay far less for input. The API accepts OpenAI and Anthropic request formats, so an existing client mostly needs a new base URL and key. Cost-sensitive teams with high-volume text or agent work, and data rules that allow processing in China, get a simple prepaid account with no plans.
- Half price outside peak hoursPeak hours are 01:00-04:00 and 06:00-10:00 UTC, Monday to Friday, excluding Chinese public holidays; all other hours cost half. Our scenario is $54 a month at peak and $27 off-peak (computed).Evidence: Pricing page (2026-10-15)
- Cache hits at 98% off, with no setupDisk caching is enabled by default for all users, and a cache hit on deepseek-flash costs $0.006 per 1M tokens against $0.30 for a miss. The guide calls it best-effort, with no guaranteed hit rate.Evidence: Pricing page; Context Caching guide (2026-10-15)
- OpenAI and Anthropic request formatsThe docs give an OpenAI-format base URL (Chat Completions and Responses) and an Anthropic-format one, and the examples use the OpenAI SDK for Python and Node.Evidence: Your First API Call; Anthropic API and Responses API guides (2026-10-15)
- A balance that does not expireThe FAQ says top-ups do not expire and unused balances can be refunded; the terms say the refund is the remaining balance after review, minus costs such as handling fees.Evidence: FAQ; Open Platform Terms of Service (2026-10-15)
- Best for
- Cost-sensitive teams running high-volume text or agent workloads with repeated prompt prefixes or flexible timing, whose data rules allow processing in China.
- Good to know
- The privacy policy says data is processed and stored in the PRC, training use is opt-out, and no SOC 2, ISO 27001, DPA or zero-retention option is stated.
How DeepSeek API compares
Against 3 other products we reviewed in LLM APIs: what you pay for with each, the five access facts, the plan limits we recorded and each review's bottom line. The note under the table says how each one was checked.
| Metric | DeepSeek API | Mistral AI API | Anthropic API (Claude) | OpenAI API |
|---|---|---|---|---|
| What you're paying for | Worth the priceA 1M-token-context model billed per token, with automatic cache hits at 98% off and half-price off-peak hours. | Worth the priceLow per-token prices on small open-weight models, with EU or US processing you can choose yourself and spending caps per workspace. | Worth the priceClaude models from Haiku 5.5 at $0.10 / $0.50 per 1M tokens up to Fable 5.1, with no training on API data and a BAA you can sign in the console. | Saves engineering timeOne API key for OpenAI's GPT-6 models and hosted tools, priced per token with published batch, flex and caching discounts. |
| Overall score | 3.2 | 7.4Best | 7.2 | 7.1 |
| Free option | Paid only | Free plan | Free trial | –Not measured |
| Card to start | –Not measured | No card needed | –Not measured | –Not measured |
| Getting access | Self-serve | Self-serve | Self-serve | Self-serve |
| Starter priceCheapest paid plan | $0.30/1M input tokens | $0.10/1M input tokens | $0.10/1M input tokens | $0.10/1M input tokens |
| Pricing model | Usage-based | Usage-based | Usage-based | Usage-based |
| Pricing page | Public prices | Public prices | Public prices | Public prices |
| How you pay | Prepaid balance (PayPal, card, Alipay, WeChat Pay); top-ups do not expire | Prepaid pay-as-you-go credits, optional auto recharge | Prepaid credits by card, expire after 1 year; invoicing through sales | Prepaid credits ($5 minimum), expire after 1 year |
| Free option | None stated; a "granted balance" is mentioned with no amount | Free mode, no card; allowance not stated publicly | One-time free credits for new users, amount not stated | No free allowance stated |
| Reference scenario (100M in + 20M out tokens/mo) | $54 on deepseek-flash at peak ($27 off-peak), computed | $12 on ministral-3b-2512, computed | $20 on claude-haiku-5-5, computed | $20 on gpt-6-luna, computed |
| Batch, cache and other discounts | No batch API; cache hits 98% off (automatic); off-peak 50% off | Batch 50% off; cached input 90% off | Batch 50% off; cache hits 90% off on Haiku 5.5 | Batch and Flex 50% off; cached input 90% off |
| Trained on API data | May be used, de-identified, unless you opt out | Unless you opt out (Free mode and paid) | No (Commercial Terms) | No, unless you opt in |
| Default retention | No number of days ("as long as you have an account") | 30 rolling days for abuse monitoring | Deleted within 30 days | Abuse logs up to 30 days |
| Zero data retention | Not stated | By approval, pay-as-you-go, stateless endpoints only | By agreement, through sales | By approval, through sales |
| SOC 2 / ISO 27001 | Not stated | Both, stated in the help center | Both, listed for the API | Both, covering the API |
| HIPAA BAA | Not stated | Not stated | Standard BAA signed in the Console, if eligible | Standard BAA accepted in the API Platform, if eligible |
| Data regions | No choice; processed and stored in the PRC | EU and US endpoints at +10% | US-only inference at 1.1x; otherwise global | 10 storage regions by approval; processing in US, EU, UAE |
| Official SDKs | None of its own; docs use the OpenAI and Anthropic SDKs | Python, TypeScript | Python, TypeScript, C#, Go, Java, PHP, Ruby | Python, TypeScript/JavaScript, Java (beta), Go (beta), C#, Ruby |
| Console SSO | Not stated | SAML on the Enterprise plan | Through sales | Custom/Enterprise billing plan |
| Bottom line | From its public pages, DeepSeek API has short, clear token prices, automatic 98%-off cache hits and half-price off-peak hours. Data is processed in China, training use is opt-out, and no compliance documents are stated. | From its public pages, the Mistral AI API has the lowest reference price in this comparison, EU or US endpoints you can pick yourself and spending caps. Training use is an opt-out, rate limits sit behind sign-in, and ZDR and SSO need approval or the Enterprise plan. | From its public pages, the Anthropic API has a complete per-token price list, discounts that combine and strong data terms, with a self-serve HIPAA BAA. Region choice is US only, zero data retention and SSO go through sales, and a few pages disagree on billing details. | From its public pages, the OpenAI API has a complete per-token price list, published batch, flex and caching discounts, and audit reports that cover the API. Stricter data controls, residency and SSO go through sales, and a few docs pages lag behind the price list. |
Evidence: DeepSeek API: facts checked 2026-10-15 from its public sign-up and pricing pages, no completed hands-on run; Mistral AI API: facts checked 2026-10-15 from its public sign-up and pricing pages, no completed hands-on run; Anthropic API (Claude): facts checked 2026-10-15 from its public sign-up and pricing pages, no completed hands-on run; OpenAI API: facts checked 2026-10-15 from its public sign-up and pricing pages, no completed hands-on run. Prices in USD as listed, excluding tax. "–" means not recorded.
Score
- Pricing clarity
- 8.0 / 10, Excellent
- per-token prices published for every model; paid online; the V4.1-Flash news post says deepseek-v4-pro requests route to Flash at Flash rates from Sept 14, 2026, while the change log and pricing page keep a separate V4 Pro price; no price stated for Files API storage
- Free tier
- 0.0 / 10, Fair
- no free tier or credits stated
- Price for 100M input + 20M output tokens
- 4.8 / 10, Fair
- $54/mo on deepseek-flash: 100M input tokens at $0.30 per 1M (cache miss, peak) plus 20M output tokens at $1.20 per 1M (peak) = $30 + $24; off-peak rates are half
- Data use and retention
- 1.5 / 10, Fair
- used for training unless you opt out; retention period not stated; no zero data retention stated
- Compliance and regions
- 0.0 / 10, Fair
- no SOC 2 stated, no regional processing option stated
- Cost controls and developer tools
- 4.8 / 10, Fair
- no batch discount, cached input 98% off, off-peak rates at half the peak rate, some rate limits published, no official SDK, structured output, compatible endpoint
Review
Review from public pages
Written from DeepSeek API's public pages (pricing, security or privacy, and docs) as read on Oct 15, 2026. We did not sign up, and no hands-on run is published yet.
What the public pages show
We read DeepSeek's API docs on Oct 15, 2026: the Models & Pricing page, the rate limit, token usage and error code pages, the guides for context caching, tool calls, JSON output, thinking mode, vision, the Files API and the Responses and Anthropic formats, the change log and the V4.1-Flash news post, and the Open Platform Terms of Service. The FAQ page is JavaScript-only, so we read its text from the script file the page loads. The English Terms of Use and privacy policy URLs served a Korean version from our location, so we read Internet Archive copies of the English pages (Oct 12 and Oct 9, 2026). We found no security or trust page. We did not sign up or call the API.
The pricing page is short and clear: two models, USD prices per 1M tokens, separate rows for cache hits and misses, peak and off-peak rates with the hours in UTC, and a concurrency limit per model. There are no plans or tiers; usage comes out of a prepaid balance, with any granted balance used first.
A few points need a question. The V4.1-Flash news post says deepseek-v4-pro requests route to Flash at Flash rates from Sept 14, 2026, while the change log and pricing page keep V4 Pro at its own price. No page states how much granted balance a new account gets. The Terms of Use name an "Improve the model for everyone" setting for opting out of training, but no API page says where it is; the privacy policy also offers an opt-out by email. Thinking mode is on by default, and the pricing page bills the total input and output tokens without a separate line for reasoning tokens.
Strengths
- Automatic cache hits at 98% off
Caching is on by default, and on deepseek-flash a cache hit costs $0.006 per 1M input tokens against $0.30 for a miss. Usage reports cache hit and miss tokens separately.
vs Google Gemini API: OpenAI API, Anthropic API (Claude), Google Gemini API and Mistral AI API price cached input on their scenario models at 90% off.
- Off-peak hours at half price
Every request outside the published peak hours is billed at half the peak rate, with no separate batch job or tier to choose.
vs Google Gemini API: OpenAI API (Flex, beta) and Google Gemini API (Flex, preview) also offer 50% off, as a separate processing tier.
- Three request formats on one key
OpenAI Chat Completions, OpenAI Responses and Anthropic Messages formats all work, and Claude model names are mapped to DeepSeek models on the Anthropic endpoint.
vs Google Gemini API: Anthropic API (Claude), Google Gemini API and Mistral AI API document an OpenAI-compatible endpoint; OpenAI API documents none in another vendor's format.
- A prepaid balance that does not expire
The FAQ says top-ups do not expire and unused balances can be refunded after review, minus handling fees.
vs Google Gemini API: Prepaid credits expire after a year on OpenAI API and Anthropic API (Claude), and after 12 months on Google Gemini API.
- 1M-token context on both models
deepseek-flash and deepseek-v4-pro both list a 1M-token context and up to 384K output tokens, with tool calls; Flash also takes images, billed as tokens.
vs Mistral AI API: Mistral AI API's scenario model, Ministral 3 3B, lists a 256k context.
Things to know
- Data is processed and stored in China
The English privacy policy says DeepSeek collects, processes and stores personal data in the People's Republic of China, and the API pages offer no region choice. The terms follow PRC law.
Workaround: The launch post links the V4.1-Flash weights on Hugging Face, for teams that can run the model elsewhere (licence not checked here).
- Training use unless you opt out
Terms of Use 4.3 let DeepSeek use de-identified inputs and outputs to develop or improve its services unless you opt out; the API terms add no API-specific rule.
Workaround: The privacy policy lists a right to opt out of training, depending on where you live, by email to [email protected].
- No compliance documents stated
No page we read states a SOC 2 report, ISO 27001, a HIPAA BAA, a DPA or zero data retention, and the privacy policy keeps data "for as long as you have an account".
Workaround: Ask [email protected] before sending personal or regulated data.
- No batch API, spend limit or token-rate numbers
Published limits are concurrency only (2,500 for Flash, 500 for Pro, per account). No spend limit or budget alert is stated; requests fail with a 402 error when the balance runs out.
Workaround: Keep the prepaid balance small and track spend with the per-key CSV usage export; higher concurrency is free on request.
Pricing reality
DeepSeek bills per 1M tokens in USD from a prepaid balance (any granted balance is used first), with no plans, seats or stated minimum spend. Every feature of a model is in its token price. Our scenario on deepseek-flash is 100 × $0.30 + 20 × $1.20 = $54 a month at peak, or $27 if all of it ran off-peak (computed). Cache hits cut input to $0.006 per 1M tokens. Costs rise with deepseek-v4-pro at $1.32 / $3.96 per 1M at peak, and with thinking mode, which is on by default and adds reasoning output. Peak hours follow the Chinese calendar, so the cheaper hours shift on Chinese public holidays. Files API storage (25 GiB per user) has no stated price, refunds deduct handling fees, and taxes are not stated. Prices can change, and changed twice in 2026.
Exit path
There is little to export: the Chat Completions and Responses endpoints keep no conversation state, and uploaded images can be listed and deleted through the Files API. Usage history can be exported per API key as CSV files. Because the API takes OpenAI and Anthropic request formats, moving means changing the base URL, key and model name. Ask for a refund of any unused balance before deleting the account, since deletion forfeits it.
Compared with
- Pick DeepSeek API when
- You want 98%-off cache hits, half-price off-peak hours and an Anthropic-format endpoint, at about the same scenario price ($54 against $55).
- Pick Google Gemini API when
- You want an ongoing free tier, a 50% batch discount and paid prompts kept out of training.
- Pick DeepSeek API when
- You want a 1M-token context, cache hits at 98% off and both OpenAI- and Anthropic-format endpoints.
- Pick Mistral AI API when
- You want a lower scenario price ($12 on Ministral 3 3B), a 50% batch discount, EU or US regional endpoints and stated SOC 2 Type II and ISO 27001.
- Pick DeepSeek API when
- You want half-price off-peak hours without a separate tier, an Anthropic-format endpoint and a balance that does not expire.
- Pick OpenAI API when
- You want a lower scenario price ($20 on GPT-6 Luna), API data kept out of training by default, SOC 2 and ISO 27001 reports and hard spend limits.
Choose it if
- Much of your input repeats, as in agent loops or long shared prompts, so 98%-off cache hits apply automatically.
- Your jobs can run outside the peak hours (01:00-04:00 and 06:00-10:00 UTC on weekdays) at half price.
- You already use the OpenAI or Anthropic SDKs and want to switch by changing the base URL and key.
- Your data rules allow processing and storage in China.
Look elsewhere if
- You need SOC 2, ISO 27001, a HIPAA BAA, a DPA or zero data retention from the vendor.
- Your data must stay in the US or the EU.
- You want API data kept out of training without having to opt out.
Bottom line
From its public pages, DeepSeek API has short, clear token prices, automatic 98%-off cache hits and half-price off-peak hours. Data is processed in China, training use is opt-out, and no compliance documents are stated.
Plan limits at a glance
| Limit | As recorded |
|---|---|
| How you pay | Prepaid balance (PayPal, card, Alipay, WeChat Pay); top-ups do not expire |
| Free option | None stated; a "granted balance" is mentioned with no amount |
| Reference scenario (100M in + 20M out tokens/mo) | $54 on deepseek-flash at peak ($27 off-peak), computed |
| Batch, cache and other discounts | No batch API; cache hits 98% off (automatic); off-peak 50% off |
| Trained on API data | May be used, de-identified, unless you opt out |
| Default retention | No number of days ("as long as you have an account") |
| Zero data retention | Not stated |
| SOC 2 / ISO 27001 | Not stated |
| HIPAA BAA | Not stated |
| Data regions | No choice; processed and stored in the PRC |
| Official SDKs | None of its own; docs use the OpenAI and Anthropic SDKs |
| Console SSO | Not stated |
Read from the vendor's public pages on Oct 15, 2026. No hands-on run is published yet. How we check
We have not published a sign-up run for DeepSeek API yet. The access facts above were read from its sign-up and pricing pages on Oct 15, 2026.
We have not published API calls for DeepSeek API yet.
DeepSeek API lists 2 plans. The cheapest paid plan, deepseek-flash (peak rate), costs $0.30/1M input tokens.
- DeepSeek APIdeepseek-flash$0.30/1M input tokens$1.20 per 1M output tokens at the peak rate; cache hits $0.006; off-peak rates are half
- DeepSeek-V4.1-Flash
- 1M context, up to 384K output
- Tool calls, JSON output, vision
- Concurrency limit 2,500
- DeepSeek APIdeepseek-v4-pro$1.32/1M input tokens$3.96 per 1M output tokens at the peak rate; cache hits $0.044; off-peak rates are half
- DeepSeek-V4-Pro-0813
- 1M context, up to 384K output
- Tool calls and JSON output (no vision)
- Concurrency limit 500
- A news post says these requests route to Flash from Sept 14, 2026; the change log and pricing page keep this price
Prices as listed, USD, excluding tax; observed 2026-10-15.
Source: api-docs.deepseek.com pricing page. Prices change; the observed date above is when our agent read them.