DeepSeek API

deepseek.com · Compute, Inference, Storage · 3 records · JSON

DeepSeek API Computepaidemailno cardDvolume-quota
DeepSeek API InferencepaidcardcardD
DeepSeek Files API Storagefails: egressemailno cardRper-object-cap

DeepSeek API Compute

paidemailno cardDvolume-quota

api-docs.deepseek.com/quick_start/pricing/ · inference API · compute/deepseek-api · JSON · compare

Free tier	No free grant on the vendor's own pricing page. deepseek-flash is USD 0.15 per 1M input tokens on a cache miss and USD 0.60 per 1M output off-peak, doubling at peak; cache hits cost USD 0.003 per 1M off-peak. The concurrency limit is 2,500 for deepseek-flash and 500 for deepseek-v4-pro, at account level, and 'There is no additional cost for capacity expansion'. Context is 1M tokens with 384K maximum output. unit none; metered on wall-clock; none
Price	none stated
Per-object cap	not applicable
Retention	not applicable
Egress	not stated
Identity	email not published
Limits	volume-quota nothing binds but the balance: 2,500 concurrent requests is the highest concurrency in this class, about eight times Cloudflare's 300 requests a minute, so the only ceiling is the prepaid balance and USD 0.15 per 1M input makes a large job cheap rather than free. source nothing binds but the balance: 2,500 concurrent requests is the highest concurrency in this class, about eight times Cloudflare's 300 requests a minute, so the only ceiling is the prepaid balance and USD 0.15 per 1M input makes a large job cheap rather than free.
Binding	volume-quota nothing binds but the balance: 2,500 concurrent requests is the highest concurrency in this class, about eight times Cloudflare's 300 requests a minute, so the only ceiling is the prepaid balance and USD 0.15 per 1M input makes a large job cheap rather than free.
Interface	unknown inferred from the tooling the provider documents; not probed
Tooling	not stated
Composes with	the off-peak window is the chain that matters: at USD 0.15 per 1M input and 2,500 concurrency a weekend batch is the cheapest job in this class and needs no free tier at all.
Reliability	no reliability event recorded
Notes	Arithmetic, labelled estimate: a 100M-input-token batch at the off-peak cache-miss rate is USD 15 and finishes in hours at 2,500 concurrent requests, where Groq's 200K TPD free lane would need 500 days for the same input. The free 5M-token signup grant that tutorials still describe does not appear on the vendor's pricing page, which prices every model and names no grant; a third-party test on 2026-08-31 reported granted_balance 0.00 and a 402 on the first call, which is reported and not my measurement. Identity on the international lane is an email address.
Evidence rung	documented
Checked	2026-10-22 (live); stale on 2026-11-21
Parsed numbers	`$.limits[0].value` limit:volume-quota 300.0 requests / minute

Evidence (3)

date	source	what it proved
2026-10-22	api-docs.deepseek.com/quick_start/pricing/	'1M INPUT TOKENS (CACHE MISS) OFF-PEAK $0.15 PEAK $0.3' and '1M OUTPUT TOKENS OFF-PEAK $0.6 PEAK $1.2' for deepseek-flash, quoted
2026-10-22	api-docs.deepseek.com/quick_start/rate_limit/	'Concurrency Limit 2500' for deepseek-flash, 'calculated at the account level', quoted
2026-10-22	api-docs.deepseek.com/quick_start/pricing/	peak hours are '01:00 - 04:00 and 06:00 - 10:00 UTC, Monday through Friday', so every other hour including weekends is off-peak, quoted

More fields (1)

{
 "wave2": {
  "class": "cheapest-paid",
  "free_tier_kind": "none",
  "shape": {
   "vcpu": null,
   "ram_gib": null,
   "gpu_class": null,
   "gpu_count": null
  }
 }
}

Unknown or withheld values (10)

$.egress.metered egress-unknown
$.egress.terms not-stated
$.identity_cost.detail process
$.interface interface-unknown
$.per_object_cap.notes process
$.per_object_cap.stated not-applicable
$.reliability_note none-observed
$.retention.notes process
$.retention.stated not-applicable
$.tooling not-stated

DeepSeek API Inference

paidcardcardD

api-docs.deepseek.com/ · inference endpoint, paid · inference/deepseek-api · JSON · compare

Free tier	no free tier observed on the documentation read unit none; metered on none
Price	none stated
Per-object cap	not applicable: an inference endpoint stores nothing for the caller
Retention	not applicable
Egress	the response is the payload; no separate egress term is published
Identity	card an API key from the platform; a card on the account
Binding	none
Interface	rest endpoint: https://api.deepseek.com
Tooling	any OpenAI-compatible client; curl against POST {base}/chat/completions
Composes with	pairs with an agent as the cheapest paid per-token route; the scarce link is the prepaid balance, not a rate
Reliability	the legacy model names deepseek-v4-flash and deepseek-v4-flash-vision-exp are retired and their requests are now served by DeepSeek-V4.1-Flash-1017 at the Flash price
Notes	The rate itself is third-party until a page rendering it is read.
Evidence rung	documented the model naming and the retirement are documented from the page read today. The per-token rate did not render on any route, so no rate is carried. verified_at is the instant its evidence URLs were fetched on 2026-10-22, read from the fetch log, not typed.
Checked	2026-10-23 (live); stale on 2026-11-22

Evidence (1)

date	source	what it proved
2026-10-22	api-docs.deepseek.com/quick_start/pricing	the served models are deepseek-flash and deepseek-v4-pro; deepseek-flash is served by DeepSeek-V4.1-Flash-1017; the legacy names are accepted but their models are retired and billed at the Flash price; OpenAI and Anthropic compatible base URLs

Unknown or withheld values (2)

$.notes partly-held
$.retention.stated not-applicable

DeepSeek Files API Storage

fails: egressemailno cardRper-object-cap

api-docs.deepseek.com/guides/files_api · Chinese LLM platform /v1/files · storage/deepseek-files-api · JSON · compare

Free tier	25 GiB per user, 10,000 files, 64 MiB per file. Permanent if `expires_after` is omitted; otherwise 1 hour – 30 days. unit GiB; metered on unstated
Price	none stated
Per-object cap	64 MiB per file (67108864 bytes)
Retention	25 GiB per user, 10,000 files, 64 MiB per file. Permanent if `expires_after` is omitted; otherwise 1 hour – 30 days
Egress	Files API itself is free; only inference tokens are billed. No egress fee documented.
Identity	email email
Limits	per-object-cap 64 MiB per file source per-object-cap 64 MiB. Upload must complete within 10 minutes. Count cap 10,000 files. `purpose` must be `user_data`. Omit `expires_after` for permanence — '少写一个参数，文件就永久留在他服务器上' (omit one parameter and the file stays on their server forever).
Binding	per-object-cap the per-object cap sets the shard size and therefore the object count; it bounds the total more tightly than any headline quota
Interface	unknown OpenAI-compatible POST /files, and an Anthropic-compatible /anthropic/v1/files; also list/retrieve/delete
Tooling	openai-python with base_url https://api.deepseek.com; curl multipart; ai-sdk deepseek provider ships a file store
Composes with	25 GiB is per user, not per project — no obvious multiplicative axis except more accounts. Image-only format restriction (JPEG/PNG/GIF/WebP, sniffed from content) makes it a picture store rather than a blob store.
Reliability	NO DOWNLOAD. The endpoint set is upload/list/retrieve-metadata/delete — there is no content-download route, and the docs describe retrieval as metadata only ('Retrieving file information returns metadata; it is not a request for the model to read the image'). 25 GiB of write-only image storage. Also: DeepSeek's platform forces phone binding on first login for some account classes (`platform.deepseek.com/mobile_verification` — 'please bind your phone number upon first login'), while international email+password signup is reported to work. Uncertain, hence medium.
Notes	25 GiB per user, 10,000 files, 64 MiB per file. Permanent if `expires_after` is omitted; otherwise 1 hour – 30 days.
Evidence rung	reported One page read moves it to documented.
Checked	2026-10-21 (inherited); stale on 2026-11-04
Parsed numbers	`$.limits[0].value` limit:per-object-cap 64.0 MiB

Evidence (6)

date	source	what it proved
2026-10-21	api-docs.deepseek.com/guides/files_api	'A single file may be at most 64 MiB'; 'expires_after[seconds] … between 3600 and 2592000. Omit both expires_after fields to keep the file permanently'; supported formats JPEG/PNG/GIF/WebP only
2026-10-21	www.datastudios.org/post/deepseek-file-uploading-supported-f	'25 GiB of stored files per user; up to 10,000 stored files'
2026-10-21	news.cnyes.com/news/id/6584205	'單個文件最大 64 MiB，每個用戶最多存 25 GiB、10000 個文件 … 不設定有效期則可以一直保留'
2026-10-21	linux.do/top?page=395	'DeepSeek 上线 Files API，每用户 25 GiB 永不过期！… 仅支持图片格式，单文件限制 64 MiB' (author @Tealina, 2026-08-21)
2026-10-21	chat-deep.ai/docs/deepseek-api-file-upload/	'Retrieving file information returns metadata; it is not a request for the model to read the image'
2026-10-21	api.deepseek.com/files	PROBE: curl -D - -> 401, Content-Length 31, body 'Authentication Fails (governor)' — live, ELB + CloudFront fronted

More fields (1)

{
 "dead_end": {
  "failure_class": "retention",
  "why": "25 GiB per user, 10,000 files, 64 MiB per file. Permanent if `expires_after` is omitted; otherwise 1 hour – 30 days.",
  "retry_when": null,
  "first_recorded": "2026-10-21"
 }
}

Unknown or withheld values (5)

$.extras.dead_end.retry_when process
$.free_tier.notes process
$.interface interface-unknown
$.per_object_cap.notes not-extracted
$.retention.notes not-extracted