Listing checkedPaidDeepSeek
Open-weight AI lab whose hosted V4 models sit behind an OpenAI-compatible API, priced from $0.14 per million input tokens on a cache miss.
- Ratings
- No member reviews yet
- Price
- Paid
- Last updated
Key takeaways
- DeepSeek is a Chinese AI lab, associated with Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd., that builds open-weight foundation models and sells hosted access to them through an OpenAI-compatible API, with a consumer web chat and mobile app on top. As of August 2026 the hosted lineup is the V4 family: DeepSeek-V4-Flash, DeepSeek-V4-Pro, and an experimental vision model, the two text models documented with 1M-token context and up to 384K output.
- The center of gravity has moved from research releases to a developer product. The first public models were DeepSeek Coder and DeepSeek-LLM in November 2023, R1 made the consumer splash in January 2025 with a claimed parity with OpenAI o1 under MIT terms, and the April 2026 V4 Preview announcement leans on coding-agent integrations, million-token context, and an Anthropic-compatible endpoint alongside the OpenAI-compatible one.
- Pricing is per-token and prepaid, not per-seat. As of August 31, 2026, V4-Flash costs $0.14 per 1M input tokens on a cache miss, $0.0028 on a cache hit, and $0.28 per 1M output tokens; V4-Pro runs $0.435, $0.003625, and $0.87 on the same axes. There is no published monthly plan, no confirmed permanent free API tier, and refunds are discretionary under the platform terms.
- Traffic estimates put deepseek.com among the most-visited AI properties: roughly 373.5M visits in June 2026 per a third-party estimate, though that figure is unverified. Nearly half of visits come from China and about 77% arrive direct.
- The biggest caveat is the 2025 security record. KrebsOnSecurity reported an exposed database found by Wiz containing chat history and API secrets in January 2025, and NowSecure findings the next month that the iOS app disabled Apple Transport Security and sent some device data unencrypted. A current, unambiguous statement on whether API prompts are used for training could not be confirmed.
What is DeepSeek?
DeepSeek develops open-weight foundation models and sells hosted access to them through an API, while also running a consumer chat experience on the web and mobile. It is a model developer first, not an interface on someone else's models: the company has released and open-sourced model families from DeepSeek-LLM and DeepSeek Coder in November 2023 through V3, R1, V3.1, V3.2, and the V4 Preview in April 2026. The homepage carries the Chinese tagline "探索未至之境", which translates roughly as "Explore the unexplored".
The hosted product, as of August 2026, is the V4 family. DeepSeek-V4-Flash is the fast, lower-cost model and DeepSeek-V4-Pro the reasoning-oriented one; both are documented with 1M-token context and up to 384K output, and both run in thinking and non-thinking modes. The company's April 2026 announcement describes Flash as 284B total and 13B active parameters, Pro as 1.6T total and 49B active, and says V4 was open-sourced. An experimental vision model, DeepSeek-V4-Flash-Vision-Exp, accepts image input under its own identifier. DeepSeek scheduled the older deepseek-chat and deepseek-reasoner endpoints for retirement on July 24, 2026.
What sits around the models is built for drop-in adoption. The API speaks the OpenAI format at api.deepseek.com and an Anthropic-compatible format at api.deepseek.com/anthropic, and the documentation lists integrations with coding and agent tools including Claude Code, OpenClaw, and OpenCode. Price is the draw: R1 listed at $2.19 per 1M output tokens in January 2025, and V4-Flash costs $0.28 today, a fraction of what DeepSeek itself charged then.
How it works
Pick a surface
Open the web chat or the mobile app for conversational use, or create an account on the DeepSeek platform for API access. The API requires a key, which you apply for after signing up.
Choose a model
API users call deepseek-v4-flash, deepseek-v4-pro, or the experimental deepseek-v4-flash-vision-exp. The stable model IDs point at updated builds, 0731 for Flash and 0813 for Pro as of August 2026, so your code doesn't change when the build does.
Send a request
Post messages to the OpenAI-compatible endpoint at https://api.deepseek.com, or use the Anthropic-compatible endpoint at https://api.deepseek.com/anthropic if your tooling speaks that format. Where supported, you can enable thinking, JSON output, tool calls, and streaming; restrictions are model-specific.
Pay by usage
Billing is prepaid. Cost is calculated from input and output tokens and deducted from your topped-up or granted balance, with granted balance consumed first. Cache-hit input tokens are priced far below cache misses, so repeated-context workloads cost less.
Features
DeepSeek-V4-Flash
The fast, lower-cost model, with thinking and non-thinking modes, 1M-token context, and up to 384K output. The company describes it as 284B total parameters with 13B active. Its documented concurrency limit is 2,500.
DeepSeek-V4-Pro
The reasoning-oriented model at a higher price, with the same documented 1M context and 384K output envelope. The April 2026 announcement puts it at 1.6T total parameters with 49B active. Concurrency is capped at 500.
DeepSeek-V4-Flash-Vision-Exp
An experimental model that accepts image input under its own model identifier. It sits alongside the two text models in the current API documentation.
Thinking mode
V4 models run in thinking or non-thinking operation, and the API documents reasoning-effort controls in integrations. This is a per-request choice rather than a separate product.
JSON output and tool calls
Structured JSON responses and function calling are listed as supported for the V4 models through the OpenAI-compatible API, though support varies by model. Chat-prefix completion and beta fill-in-the-middle completion are also available, FIM in non-thinking mode only.
Context Caching
Pricing distinguishes cache-hit from cache-miss input tokens, and a hit costs a small fraction of a miss: $0.0028 against $0.14 per 1M tokens on V4-Flash. Workloads that resend the same context, such as agents and long documents, benefit most.
Open weights
DeepSeek publishes the checkpoints themselves, on top of the hosted endpoints. The DeepSeek-LLM 7B and 67B base and chat models were published for research and commercial use under stated license terms in November 2023, R1 shipped under MIT terms in January 2025, and the company says V4 was open-sourced.
What people use it for
- Conversational assistants and customer-support chatbots.
- Code completion, fill-in-the-middle completion, debugging, and agentic coding.
- Long-context document summarization and question answering.
- Mathematical and general reasoning tasks.
- Structured JSON extraction and tool-using workflows.
- Translation, drafting, rewriting, and search-style retrieval.
It is not a safe default for sensitive production workloads: security researchers reported serious iOS transport, device-data, and exposed-database problems in 2025, and DeepSeek's current position on whether API prompts are used for training could not be confirmed.
Who it’s for
Pricing
Official pricingIndividual plans
DeepSeek-V4-Flash (API)
$0.14
per 1M input tokens
Cache-miss input; cache hits $0.0028 and output $0.28 per 1M tokens, 1M context, up to 384K output
DeepSeek-V4-Pro (API)
$0.435
per 1M input tokens
Reasoning-oriented; cache hits $0.003625 and output $0.87 per 1M tokens, 1M context, up to 384K output
There's no plan menu: as of August 31, 2026 the official pricing is per-token, prepaid, and model-specific. V4-Flash costs $0.14 per 1M input tokens on a cache miss, $0.0028 on a cache hit, and $0.28 per 1M output tokens; V4-Pro costs $0.435, $0.003625, and $0.87 respectively. You spend the balance down as you use the service, and DeepSeek may suspend service when it runs out; the company also reserves the right to change fees, and it has, materially: R1 launched in January 2025 at $2.19 per 1M output tokens and V3 in December 2024 at $1.10. The refund terms deserve a careful read. Refund requests are reviewed rather than granted, only the remaining unspent balance is refunded if approved, it comes as one lump sum after possible handling costs, partial refunds are not supported, and spent amounts are never refunded. No permanent free API tier, top-up pack price, credit expiry rule, or discount schedule is confirmed on official pages; a third-party guide from August 2026 claims free web chat and a 5M-token new-account grant valid for 30 days, but treat that as unverified until you see it in the account UI.
Traffic & reach
Third-party estimates, not official numbers, and in DeepSeek's case unverified ones: the roughly 373.5M visits estimated for June 2026 could not be independently confirmed. Taken at face value the shape says as much as the size. Nearly half of visits come from China, about 77% arrive direct, and the June figure is down around 13% from a May peak near 430M after climbing through the spring.
Visits, June 2026
373.5M
▼ 13% month over month
Nov '25Jun '26
Top countries
- China48%
- Russia9.9%
- United States5.7%
- Hong Kong SAR China3.1%
- Brazil3.0%
Traffic sources
- Direct77%
- Organic search14%
- Referrals3.3%
- Organic social2.7%
- AI assistants2.6%
- Visits, last 3 months
- 1215.1M
- Avg. visit
- 5m 14s
- Pages per visit
- 3.3
- Bounce rate
- 38.1%
- Brand searches / mo
- 16.9M
User reviews
Write a reviewNo reviews yet
Used DeepSeek on something real? Write down what you made, what it cost and what the trade-off was. The next person can decide in about a minute from that.
Be the first to reviewWhat to know before you commit
Priced to undercut. The clearest reason people pick DeepSeek is cost. Current V4 pricing sits far below the company's own historical V3 and R1 rates for high-volume token use, cache-hit input most of all at $0.0028 per 1M tokens. That makes it attractive for experimentation and price-sensitive inference.
The open-weight strategy is real. R1 shipped with a technical report and MIT licensing language, the earlier DeepSeek-LLM repository provides checkpoints and training information, and the company says V4 was open-sourced. The reasoning claims are another matter: R1's claimed parity with OpenAI o1 and V4's coding and context claims are the company's own, not independently benchmarked in what we reviewed.
The 2025 security record is serious. KrebsOnSecurity reported that Wiz found an unauthenticated DeepSeek database in January 2025 exposing chat history, logs, API secrets, and operational data, and that NowSecure researchers found the iOS app had disabled Apple Transport Security, sent some device information in plaintext, used deprecated 3DES with a hard-coded key, and communicated with ByteDance's Volcengine. The reporting we reviewed does not document a complete remediation statement from DeepSeek.
Expect availability wobbles. TechRadar reported a large cyberattack in January 2025 that led DeepSeek to limit registrations. A later monitoring report claims five-plus outages through February 2026 and 97.8% uptime, though that figure comes from a secondary tracker whose methodology is not independently verified.
No verified ratings anywhere. We couldn't find a reliable aggregate rating with a review count on G2, Capterra, Trustpilot, Product Hunt, or the app stores, and no attributable user quotes met a publishable standard. The reception picture here is built from pricing pages, official announcements, and press reporting, not review-site scores.
Alternatives to DeepSeek
OpenAIThe company behind ChatGPT, the Responses API and the GPT-5.6 models: consumer AI apps, developer APIs and enterprise workspaces with a permanent free tier.
Hostinger Managed OpenClawManaged hosting for OpenClaw: a private AI agent in an isolated Docker container, paired to WhatsApp or Telegram and running 24/7 with no server to maintain.
ClaudeAnthropic's AI assistant for writing, coding, research and analysis, with a permanent free tier and paid plans from $20 a month.
JuicyChat AIAdult character chat with community-made and custom personas, plus coin-metered images and voice on web, iOS and Android.
LovableDescribe a web app in chat and Lovable writes real full-stack code, provisions the database, auth and hosting, and syncs the repo to GitHub or GitLab.
CrushOn.AIAdult character roleplay and AI companion chat with user-created characters, plan-based memory up to 24K and a permanent free tier on free models.
Frequently asked questions
Company & links
- Company
- DeepSeek
- Founded
- 2023
- Leadership
- Liang Wenfeng is described as DeepSeek's CEO and is also the founder and leader of the High-Flyer hedge fund. No fuller executive roster has been confirmed beyond him.
- Funding
- DeepSeek is described as owned and funded by High-Flyer rather than venture-backed in the usual sense, and no primary-source funding announcements exist. Wikipedia's entry claims a May 2026 Series A at a $52 billion post-money valuation, but no independent financial publication or filing corroborates it, so treat both the round and the valuation as unverified.
- Scale
- Not documented. No verified headcount, headquarters street address, revenue, or run-rate figure was available as of August 2026; the homepage showed active recruiting when we checked.
- Contact
- service@deepseek.com
Try DeepSeek on work you already know
A task you’ve done before is the fastest way to judge whether DeepSeek fits your workflow and budget. Check the current plan terms on the official site before you subscribe.