DeepSeek

There are two things to consider: whether you need a chat without payment or workarounds, and what you plan to upload to it.

Who it is for. Anyone who wants a strong AI for free and without hassle: the first question can be asked without signing up.

Who else. For developers and companies that cannot send data abroad: V4 weights are open under MIT, so you can deploy the same model on your own hardware.

Who it is not for. For people who work with other people's data and contracts: according to the vendor's policy, conversations are stored in China (Feb 10, 2026).

Suitable: section 3 gets you to your first question in three steps. If you're choosing between it and ChatGPT: the choice is covered in section 7.

What vibecoding isStarting from scratch: how to set tasks for agents with no coding experience

Questions, search and files in one window, and its own V4.1 Flash model under the hood.

Think or answer right away. Since Sep 10, 2026 there are no Instant, Expert and Vision modes in the chat: the model itself decides how long to think about a question. The DeepThink button under the input field turns on long reasoning: in the official storefront screenshot the model "thought for 52 seconds" about a sums-of-squares problem. The expert mode with V4 Pro is gone, and App Store reviews ask for it back.

Search and images. The Search button pulls in sites and shows how many pages it found and read. Photos, screenshots and documents attach with the attachment button, a separate mode for images is no longer needed. In the storefront screenshot a 1.2 MB text file with a novel is dropped into the chat, and the model retells the plot from it.

Under the hood. Its own V4.1 Flash model with a context of a million tokens, several thick books in one conversation. It has 552 billion parameters, but only 16 billion work on an answer, so it is fast and cheap. The weights are open under the MIT license, and other clouds run it, while the former flagship V4 Pro with 1.6 trillion parameters stays in the API.

For developers. On Aug 13, 2026 the vendor released DeepSeek Harness, its own coding agent, in preview for developers; it has a separate page in the catalog.

What was actually done. On Habr, the M.Video team gave five models the longest-palindrome-in-a-string task (Nov 26, 2025): DeepSeek solved it 10 out of 10, but the code almost matched Grok's answer, and the author suspected a memorized template. We have not run DeepSeek on that task ourselves; the numbers come from official materials and other people's tests.

Step 1. Ask your first question without registering; the basic chat works anonymously.

Step 2. Create an account for history and files: sign in with email and password, Google, or Apple.

Step 3. Ask your question. For a hard task turn on DeepThink under the input field, for fresh data from the web the search button.

Apps. iPhone: the app is in the App Store, version 2.5.2 from Sep 20, 2026. Android: the official package is in Google Play with 50 million downloads worldwide, and without Google Play the official APK file sits on the DeepSeek download page. DeepSeek has no desktop programs, only the browser.

Stability. There are outages. September reviews repeat "server is busy" and "editing too often", and the second one means a limit on editing a question and regenerating an answer.

Steps checked Sep 26, 2026 against the App Store and Google Play storefronts, the chat.deepseek.com sign-in screens verified Sep 5, 2026

The model is below the top tier, but its audience is among the leaders.

Artificial Analysis index. V4.1 Flash scores 39 points (Sep 26, 2026), leader Claude Opus 5.5 scores 58, and the openGLM-5.3 and Kimi K3 score 45 and 44. But an index task costs DeepSeek $0.27 versus $2 for Kimi K3 and GLM-5.3, seven times cheaper. Not first on points, but several times cheaper than its neighbors.

Vendor claims. In the V4.1 Flash release log (Sep 10, 2026) Terminal Bench 2.1 stands at 90.6 against 87.9 for V4 Pro, and the vendor writes that the new model beats the former flagship. Some of the sets are internal, so treat individual numbers as self-reported, but an independent index confirmed the overall conclusion: 39 points against 36 for V4 Pro.

Audience. Similarweb counted 315M visits to chat.deepseek.com in July 2026 and fifth place among the world's AI chats.

Its place among the assistants in our catalog is shown under the card title, and the full tool ranking is on the shared page.

The chat is free in full, only developers pay, and since Sep 10, 2026 the new model costs them less than the old one.

Free entry. The chat has no paid tier: the site lists only API pricing, and there are no in-app purchases in the App Store (Sep 26, 2026). The vendor does not publish limits, at peak hours the chat replies with a "server is busy" error, and since September the app has restricted editing a question and regenerating an answer.

API by hour. The price per million tokens depends on the time of day: on weekdays the peak runs from 01:00 to 04:00 and from 06:00 to 10:00 UTC. V4.1 Flash charges $0.15 for input and $0.60 for output off peak and twice that at peak, V4 Pro $0.66 and $1.98 off peak. Repeating context already sent costs $0.003 on V4.1 Flash, fifty times cheaper.

What changed. Since Aug 16, 2026 V4 Flash output has gone up from $0.28 to $0.66 off peak, and Habr summed it up: the era of an almost free DeepSeek API is over (Aug 14, 2026). On Sep 10, 2026 V4.1 Flash came out and prices went back down: off-peak input fell from $0.22 to $0.15, output from $0.66 to $0.60.

Three questions: where correspondence is stored, what the model stays silent about, and when it makes mistakes.

Data in China. Under the privacy policy dated Feb 10, 2026 the vendor collects, processes and stores data in China, and the company answering for it is based in Hangzhou. Your conversations feed model training; opting out is declared a user right, but the policy does not say where to switch it on. The rule is simple: do not upload contracts, personal data or work files into the chat. If you need the same model without sending data abroad, run the open weights on your own hardware.

Censorship. In an independent ellamind test on 168 topics sensitive for China, DeepSeek V3.2 answered honestly in only 19% of cases, while Kimi K2.5 did in 99% (February 2026). There is no such measurement for V4 and V4.1.

Incidents. In January 2025, Wiz researchers found an exposed DeepSeek database containing chat histories and API keys; the vendor closed it after notification. On Jul 21, 2026, search engines indexed conversations users had themselves shared via public links; private chats were unaffected. On Feb 23, 2026, Anthropic accused DeepSeek, Moonshot, and MiniMax of training on Claude responses through thousands of fake accounts; the companies did not respond to CNBC's request.

Government restrictions. Under the NDAA (signed Dec 18, 2025) the US Department of Defense has to remove DeepSeek models from its systems and ban them for its contractors, Italy restricted processing of its users' data from Jan 30, 2025, and government agencies in Australia and South Korea banned the service on work devices in February 2025.

The choice is not "who is smarter," but "how much access costs and what you pay for it."

Access. DeepSeek opens with no signup and no card. ChatGPT's Plus subscription costs $20 a month.

Smarts. On the Artificial Analysis index GPT-6 Astra scores 53 against 39 for V4.1 Flash (Sep 26, 2026). By AA's measurement DeepSeek is fast but verbose, the index task costs it $0.27 against $3.26 for Astra, and the same model is available in open weights.

Data. DeepSeek stores conversations in China, while ChatGPT stores them in the US. For work documents, both services are foreign; the only difference is jurisdiction.

Who else is nearby. Kimi handles long documents and a swarm of agents.

In three years the lab inside a hedge fund went from the open V3 model to a chat used around the world.