DeepSeek Review 2026: A Free Chatbot, MIT-Licensed Weights, and an API Priced at Cents -- With Fine Print You Should Read First
DeepSeek is the AI lab that keeps embarrassing much bigger budgets. Its newest model, V4.1 Flash (released Sept 10), costs $0.15 per million input tokens and $0.60 per million output tokens off-peak, ships with open weights under the MIT license, and, on many of the tests DeepSeek published, scores ahead of its own V4 Pro. The chat app is free. But the product you would try today is not the one in most older reviews: V4 Pro was announced as being phased out four days after V4.1 Flash launched, then kept alive after users objected; DeepSeek has launched an open-source agent called Harness with a desktop app; and the privacy policy still says your data is stored in the People's Republic of China. This review checks each of those against DeepSeek's own docs, then adds what real users say.
| Operator | Hangzhou DeepSeek AI Co., Ltd. (China) |
| Chat app | Free (web, iOS, Android) |
| Cheapest API model | $0.15 / $0.60 per 1M in/out, off-peak |
| Open weights | MIT (V4.1 Flash on Hugging Face) |
| Trustpilot record | 2.3/5, 155 reviews |
The best value in AI right now, if you accept that you are dealing with a fast-moving Chinese lab, not a polished consumer brand.
"DeepSeek's strength is not that it wins every benchmark. It does not. Its own published table has it behind GPT-5.6 Sol and Claude Opus 5 on several hard tests. The strength is that it gets close to them for a fraction of the cost, gives the weights away under an MIT license, and keeps doing it every few weeks. The weakness is the other side of the same coin: models get retired and swapped quickly, the chat app has real quirks, and the consumer service stores your data in China."
My honest read: use the free chat for non-sensitive questions, use the API when cost per token is the point, and self-host or use a third-party host if you cannot let data leave your jurisdiction. Do not paste client secrets, medical data or unreleased code into the consumer app. If you need one polished all-purpose assistant with support and predictable behaviour, the alternatives section below is where to look.
A free chat app, a pay-as-you-go API, open model weights, and (new) an open-source agent called Harness.
As of this check DeepSeek has four doors. The chat app (chat.deepseek.com, iOS, Android) is free and offers an Instant mode and an Expert mode plus DeepThink and web search toggles. The API platform sells the models by the token and speaks both OpenAI and Anthropic formats. The open weights are published on Hugging Face under the MIT license, including DeepSeek-V4.1-Flash. And DeepSeek Harness (dsh) is an MIT-licensed agent harness, now with desktop apps for macOS (Apple silicon) and Windows, still marked a developer preview.
One piece of context up front: I could find no official DeepSeek page describing its funding or investors, so this review does not make claims about who owns the company beyond what its privacy policy states: the service is provided and controlled by Hangzhou DeepSeek Artificial Intelligence Co., Ltd., registered in China.
Huge among developers and in China, a small slice of Western chatbot traffic.
deepseek-ai/deepseek-harness, created Aug 13, 2026; read from GitHub's API Oct 3. Stars measure interest, not users.
Hugging Face model counter on the official repo, Oct 3. Downloads are not users either.
Google Play listing for the official app, updated Sep 30. A Play bucket, not active users.
Similarweb, August 2026, via search-aggregated summaries (not re-checked at source). About 6% of chatgpt.com's visits.
| Date | Event | Source |
|---|---|---|
| 2026-02-10 | Current privacy policy version published | DeepSeek privacy policy |
| 2026-04-24 | V4 Pro (1.6T total / 49B active) and V4 Flash (284B / 13B) released, open weights, 1M context | DeepSeek API docs news |
| 2026-07-31 | V4-Flash official API update in public beta | DeepSeek change log |
| 2026-08-13 | V4 Pro GA; low / high / max thinking effort; Responses API; Harness repo created the same day | Change log; GitHub |
| 2026-08-16 | Peak / off-peak API pricing starts (off-peak is half) | DeepSeek API docs news |
| 2026-08-21 | V4-Flash-Vision-Exp and a free Files API | DeepSeek API docs news |
| 2026-09-10 | V4.1 Flash released, V4 Pro phase-out announced, prices cut | DeepSeek news |
| 2026-09-14 | Planned V4 Pro re-route; DeepSeek later said it would keep serving V4 Pro | Change log; pricing page |
| 2026-09-30 | Harness desktop apps for macOS and Windows (developer preview) | deepseek.com/en/download; press |
Developers love the price and speed. General users complain about language drift, censorship and stability.
2.3 out of 5; 23% five-star, 55% one-star. Praise: "it's absolutely blown my mind" (Aug 19). Complaints: "It keeps responding in Chinese. I keep telling it not to" (Aug 9) and "Inconsistent, low quality. There are much better options" (Aug 8).
"Robust free tool, but held back by usability issues": the reviewer lists poor context retention in long chats, "over-censorship" on China-related topics with hard stops mid-answer, and the model reverting to Chinese when the chat started in English. Google Play's average is lower (2.3), with a mix of delighted and angry reviews.
Mostly positive on value: "a really strong model and the fact that they reduced prices at the same time makes it an awesome backup model." Criticism is specific: V4 Pro customers being moved without a long deprecation window, and one tester who found GLM 5.3 turned up far more vulnerabilities than DeepSeek in a kernel-audit test. Another said it "becomes really good if you provide custom tools."
Chat is free. The API is the product you pay for, and it is priced per million tokens with a half-price off-peak window.
Chat app
Web, iOS and Android.
- Instant and Expert modes
- DeepThink, web search, file upload
- No published usage cap I could find
API: deepseek-flash
V4.1 Flash. $0.60 per 1M output, off-peak.
- Peak: $0.30 in / $1.20 out
- Cache hit: $0.003 off-peak
- 1M context, vision, tools
deepseek-v4-pro
V4-Pro-0813. $1.98 per 1M output, off-peak.
- Peak: $1.32 in / $3.96 out
- No vision support
- 500 concurrency limit
Open weights
Download and run it yourself.
- No per-token fee to DeepSeek
- You pay for hardware
- Data stays where you host it
| Per 1M tokens (USD) | deepseek-flash, off-peak | deepseek-flash, peak | deepseek-v4-pro, off-peak | deepseek-v4-pro, peak |
|---|---|---|---|---|
| Input, cache hit | $0.003 | $0.006 | $0.022 | $0.044 |
| Input, cache miss | $0.15 | $0.30 | $0.66 | $1.32 |
| Output | $0.60 | $1.20 | $1.98 | $3.96 |
| Context / max output | 1M tokens context; 384K max output (both models) | |||
| Concurrency limit | 2,500 | 500 | ||
Pick by what data you are sending and what you are building.
The free chat app. Check answers; expect an occasional language slip.
The API with deepseek-flash, run heavy jobs off-peak.
Not the consumer app. Self-host the MIT weights or use a host you trust, and get your security team to sign off.
DeepSeek Harness desktop, with the developer-preview warnings in mind.
Look at ChatGPT, Claude or Gemini in Alternatives.
Perplexity is built for it; DeepSeek's search toggle is not its strength to my knowledge.
A 552B-parameter mixture-of-experts model with only 8B active parameters on input and 16B on output, built to be cheap in agent loops.
DeepSeek's Sept 10 post calls V4.1-Flash the smallest model in its new architecture family and gives the headline specs: a 552B-parameter MoE, a new causal encoder-decoder design with 8B active parameters for input and 16B for output, native visual understanding, and a KV cache that needs a quarter of the HBM and an eighth of the SSD of the previous generation. DeepSeek says cache-hit charges are a large share of agent costs, which is why compressing the cache cut prices. It also says tests by multiple parties put V4.1 Flash ahead of V4 Pro on performance, cost, speed and total runtime. The weights are on Hugging Face under the MIT license (the repo's tag is license:mit; it was created Sept 10 and updated Oct 1).
| Benchmark (DeepSeek's table) | V4.1 Flash | V4 Pro 0813 | GPT 5.6-Sol | Claude Opus 5 |
|---|---|---|---|---|
| GPQA Diamond | 90.9 | 92.4 | 94.1 | 93.4 |
| HLE (no tools) | 36.8 | 42.7* | 44.5 | 56.3 |
| Terminal-Bench 3.0 | 30.0 | 11.8 | 34.4 | 43.3 |
| DeepSWE v1.1 | 74.2 | 62.7 | 73.0 | 74.0 |
| CyberGym | 88.1 | 83.3 | 84.5 | - |
| ExploitGym | 15.3 | 5.4 | 33.7 | 22.1 |
- The price cut is real and checkable on the live pricing page.
- HN testers independently report speed (300-400 tokens per second at times) and good results as a cheap fallback model.
- Open weights mean you can inspect, fine-tune or host it yourself.
- Benchmarks are DeepSeek's own and selected by DeepSeek.
- Independent testers disagree: one HN commenter found Flash "wildly bad at doing as asked" in code, another that it needs custom tools to shine.
- The model grew from 284B to 552B parameters, so local hosting got harder (HN calculated hundreds of GB of memory).
Choose a mode, let it reason, then verify -- and decide up front which data is allowed in the box.
Two workflow details that change results. The thinking toggle (DeepThink in chat, thinking in the API) switches the same model between non-thinking and thinking, and the API accepts three effort levels, low, high and max, per the Aug 13 notes. And because peak hours are weekday mornings in UTC, scheduling bulk jobs after 10:00 UTC or at weekends halves the bill.
An open-source agent that runs on your machine, built on a "everything is a plugin" design.
DeepSeek Harness (dsh) is MIT-licensed, written in TypeScript, and its GitHub repo passed 242,000 stars within eight weeks of creation (Oct 3 API read). The README calls it a developer preview, warns that "there will be compatibility-breaking changes", and tells you to read its safety notice before running it. You can start it with npx @deepseek-ai/dsh web or use the desktop apps announced around Sept 30 for macOS (Apple silicon) and Windows. DeepSeek's page shows sessions, plugins, automation, a workspace and a "Creator mode" in which you ask the agent to write a plugin, plus experimental plugins for agent teams, scheduled tasks and voice input. A Hacker News user wrote that it "has particularly good observability" of every prompt and tool call, another that its append-only context avoids cache-invalidation bugs; a third said the thread "feels astroturfy", so weigh the enthusiasm accordingly.
- Everyday file and document chores plus coding, with any model you configure (DeepSeek or a custom API endpoint).
- Plugins and a free, MIT-licensed core you can audit.
- I have not run it. Press coverage says it can execute model-generated commands with access to your files and credentials, and the project's own safety notice is the thing to read first.
- Sessions using DeepSeek's official models generate logs of inputs and outputs (per press coverage of the release).
It speaks OpenAI and Anthropic formats, so most coding tools can use it as a backend.
DeepSeek's docs expose two base URLs, https://api.deepseek.com (OpenAI format) and https://api.deepseek.com/anthropic (Anthropic format), and since Aug 13 the OpenAI Responses API natively, adapted for Codex with a one-click script. Its Agent Integrations guide lists DeepSeek Harness, Claude Code, Codex, OpenCode, OpenClaw, Hermes, Reasonix, WorkBuddy/CodeBuddy and Qoder. The Sept 10 post names WorkBuddy (including CodeBuddy) and OpenCode as official partners that fully support V4.1 Flash. For Claude-side workflows see the Claude Code setup guide; for OpenAI's coding agent see the Codex review.
| Integration | How it works | Source |
|---|---|---|
| OpenAI Chat Completions format | Change base_url and model name in any OpenAI SDK | Docs, Your First API Call |
| Anthropic format | base_url api.deepseek.com/anthropic; Claude Code supported | Docs, Agent Integrations |
| OpenAI Responses API | Native since Aug 13; Codex one-click configuration | Change Log |
| Vision and Files API | Images up to 384 tokens each at Flash price; uploads free | Aug 21 post |
| Other agent tools | OpenCode, OpenClaw, Hermes, Reasonix, Qoder, WorkBuddy/CodeBuddy | Docs sidebar |
| MCP | Not described on the pages I read; check Harness docs and your client | - |
Output tokens are where agents spend, and DeepSeek's are a small fraction of the US labs' list prices.
| What you're doing | What you pay |
|---|---|
| Chatting in the DeepSeek app | $0 |
| 1M input + 1M output tokens on deepseek-flash, off-peak, cache miss | $0.15 + $0.60 = $0.75 |
| Same at peak | $0.30 + $1.20 = $1.50 |
| Same on deepseek-v4-pro, off-peak | $0.66 + $1.98 = $2.64 |
| Reference: OpenAI GPT-6 Astra API (list) | $10 input + $50 output per 1M (OpenAI, Sep 3) |
| Reference: OpenAI GPT-6.1 Sol API (list) | $2 input + $10 output per 1M (OpenAI, Sep 29) |
| Cost lever | Effect | Source |
|---|---|---|
| Off-peak scheduling | 50% of peak rates | Pricing page |
| Cache hits | $0.003 vs $0.15 per 1M input on Flash, off-peak | Pricing page |
| Smaller KV cache | DeepSeek says it cuts the large cache-hit share of agent bills | Sept 10 post |
| Granted balance first | Granted credit is spent before topped-up balance | Pricing page |
The consumer service stores your data in China and may use it to train models.
| Control | What the privacy policy (updated Feb 10, 2026) says |
|---|---|
| Who runs it | Hangzhou DeepSeek Artificial Intelligence Co., Ltd., registered in China |
| What it collects | Account data, text/voice input, prompts, uploaded files, chat history, IP address, device identifiers |
| Where data is stored | "we directly collect, process and store your Personal Data in People's Republic of China" |
| Use for training | Used "to train and improve our technology, such as our machine learning models"; I found no consumer opt-out in the text I read |
| Retention | As long as you have an account, plus legal and legitimate-interest retention |
| API developers | Rules for end users of apps built on the API are the developer's responsibility, not covered by this policy |
| Rights | Access, deletion, objection and complaint rights listed per local law |
The risk is not only theoretical for institutions. Governments have restricted it: in 2025 several US states, Australia and Taiwan banned it on government devices (press reports), and a "No DeepSeek on Government Devices Act" (H.R.1121) was introduced in the US House. Wikipedia also records a 2025 data-exposure incident. I did not test DeepSeek's security myself. The practical rule: treat the hosted service as you would any foreign cloud app with broad collection and training rights.
Best for developers and cost-sensitive builders; weaker for regulated work and people who want a polished consumer assistant.
| Situation | Why DeepSeek fits | Likely route |
|---|---|---|
| Student or curious user | Free, strong reasoning, no card needed | Free chat (non-sensitive) |
| Developer running agents | Very low output price, 1M context, OpenAI/Anthropic-compatible | API, deepseek-flash |
| Team with GPUs and privacy needs | MIT weights, self-hostable | Self-host or trusted third-party host |
| Tinkerer wanting a local agent | Open-source Harness with plugins | Harness desktop or npx |
| Regulated industry, government | Not recommended on the hosted service | Self-host only, with review |
| Someone who wants cited web answers | Not its sharpest edge; compare Perplexity | Free to test, then decide |
Five real gaps, not manufactured ones.
deepseek-chat and deepseek-reasoner were scheduled for retirement on Jul 24; V4 Flash and the Vision-Exp preview were retired on Sept 10; V4 Pro was set to be re-routed to Flash on Sept 14 and then, after user feedback, kept. HN users complained about being moved without a longer deprecation window. Pin model names and keep tests.
The policy says collection, processing and storage happen in the PRC, and inputs may train models. That alone rules the hosted service out for many employers.
Trustpilot and App Store reviewers report replies drifting into Chinese mid-chat, weak context retention and hard stops on China-related topics. Google Play averages 2.3.
The Sept 10 post says V4 Pro requests route to Flash from Sept 14, while the Change Log and pricing page list V4 Pro as still served. I trusted the pricing page; confirm in your own account.
DeepSeek's own table has it behind GPT 5.6-Sol and Opus 5 on HLE, ExploitGym and Terminal-Bench 3.0, and independent HN testers saw mixed results.
The official status page shows 99.69% uptime for V4.1 Flash API and 99.84% for chat over Jul-Oct 2026; an HN post on Oct 1 flagged an outage. Fine for hobbyists, a planning item for production.
DeepSeek is the cheap, open one. Others win on polish, privacy or citations.
| If you mostly need... | Compare DeepSeek with... |
|---|---|
| The most complete general assistant with images, voice, agents | ChatGPT |
| Careful long-form writing and coding | Claude |
| Google Workspace integration and a big free tier | Google Gemini |
| Cited, search-first answers | Perplexity |
| Running open models locally with your own data | Ollama and Open WebUI |
| An editor-native coding assistant | GitHub Copilot |
| A chatbot tied to real-time X data, or Microsoft 365 integration | xAI's Grok (grok.com) and Microsoft Copilot (copilot.microsoft.com) |
| Entry cost to use it | DeepSeek | ChatGPT | Self-hosted open weights |
|---|---|---|---|
| Free chat | Yes, app and web | Yes, Free plan (ads possible) | No chat UI included; run your own |
| Cheapest paid consumer plan | None found | Go at $8/mo (US) | Hardware cost |
| API list price, small model | $0.15 / $0.60 per 1M (off-peak) | See OpenAI's pricing | No per-token fee |
No affiliate relationship shapes this review.
JAVIS has no publisher affiliate program with DeepSeek, and DeepSeek has no consumer plan to sell. Every CTA in this article points to the official product. This is a VERIFIED EDITORIAL REVIEW: I did not hands-on test DeepSeek.
Go to the official chat.
Open DeepSeek chatRead the alternative that matches your real question.
Choose the next articleDeepSeek sits between the big-lab assistants and the open-source local stack.
All six cards link to reviews that are already live on JAVIS.
Card images are official assets from each product's own site or from the product's already-published JAVIS page.
Questions people are actually searching right now
Is DeepSeek free?
The chat app on web, iOS and Android is free, and I found no paid consumer plan on official pages. The API is pay-as-you-go: deepseek-flash costs $0.15 per 1M input and $0.60 per 1M output tokens off-peak, double at peak.
What is the latest DeepSeek model?
DeepSeek-V4.1-Flash, released Sept 10, 2026: a 552B-parameter MoE with native vision, available via the API as deepseek-flash and as MIT-licensed weights. DeepSeek has hinted that a V4.1 Pro is to come, but no date was given in what I read.
What happened to DeepSeek V4 Pro?
On Sept 10 DeepSeek said it was phasing V4 Pro out and re-routing requests to V4.1 Flash from Sept 14. Its Change Log later says it will keep providing V4 Pro with billing unchanged, and the pricing page still lists it.
Is DeepSeek safe to use?
It depends on the data. The privacy policy says data is stored in China and may be used for training; several governments restrict it on official devices. Fine for casual non-sensitive use; for confidential data, self-host the open weights or avoid it.
Is DeepSeek open source?
The model weights for V4 and V4.1 Flash are published under the MIT license, and the Harness agent is MIT-licensed. The training data and the hosted service are not open.
Can I run DeepSeek locally?
Yes in principle. V4.1 Flash is much larger than its predecessor (552B parameters), so HN users estimate it needs several hundred gigabytes of memory; smaller community quantisations may exist but I did not verify any.
What is DeepSeek Harness?
An open-source, MIT-licensed agent harness with plugins and desktop apps for macOS (Apple silicon) and Windows. It is a developer preview with compatibility-breaking changes expected, and it can act on your files, so read its safety notice.
What are the best alternatives?
ChatGPT for breadth, Claude for writing and code, Gemini for Google integration, Perplexity for cited research, and Ollama for local use. See Alternatives above.
Where this came from, and when it was checked.
Pricing and models: api-docs.deepseek.com/quick_start/pricing; Change Log (api-docs.deepseek.com/updates); news posts of Apr 24, Aug 13, Aug 21 and Sept 10, 2026; checked Oct 3, 2026.
Product pages: deepseek.com/en, /en/download, /en/harness, /en/transparency, /en/news/deepseek-v4-1-flash; status.deepseek.com (uptime Jul-Oct 2026).
Open source: GitHub data for deepseek-ai/deepseek-harness (MIT, ~242.7K stars, created Aug 13) and the repo README; Hugging Face API for deepseek-ai/DeepSeek-V4.1-Flash (license:mit, ~788K downloads).
Privacy and data: DeepSeek privacy policy (last updated Feb 10, 2026), cdn.deepseek.com.
Market: Google Play and Apple App Store listings (installs, ratings); Similarweb and QuestMobile figures via search-aggregated summaries, not re-verified at source; GitHub and Hugging Face counters read directly.
Real users: Trustpilot deepseek.com (2.3/5, 155 reviews, read Oct 3); App Store review text; Hacker News threads 49639090, 49624603, 49725800, 49929489 and the V4 Pro discontinuation email post 49639667 on Hacker News. Reddit could not be read directly.
Government restrictions: press and congress.gov bill listing (H.R.1121); 2025 events, not re-verified against each government's own notice.
Harness safety caveats: the repo README (developer preview) and press coverage of the Sept 30 desktop release (runtimewire.com); the SAFETY.md file itself was not read.
Official video: none embedded. deepseek.com links to no YouTube channel, and I could not confirm any channel as DeepSeek-owned.
Fit Score (74/100): editorial composite of five criteria scored out of 20: value for money 19, capability and openness 17, product breadth and polish 13, privacy and governance 11, user sentiment and stability 14. A judgement, not a benchmark.
