Skip to content
AgentsIndex

DeepSeek

AI model platform for chat, mobile, and API workflows.

Best for
Developers and teams who want chat access plus API-ready model integration.

What we checked

Pages last read
Facts checked
1 fact

From 1 source, read 29 October 2026

AI answers
Too few read

58 buyer-question answers read, November 2026

At a glance

For developers

Works with
GitHub

DeepSeek in our research

All research

Study

AI Tool Pricing Index, November 2026

Pricing

Usage-based

43 of the 361 tools we checked bill by usage.

Edition price, checked 29 October 2026.

Pricing pages of the AI tools, November 2026 editionas of n = 361 tools checked

What is DeepSeek?

DeepSeek is an AI model platform for developers and teams that moves from free chat to API integration across web, mobile app, and docs-backed workflows. It centers on API access, GitHub integration, and a mobile app, while its research pages surface model families including DeepSeek R1, V3, Coder V2, VL, Math, and LLM. The site also shows DeepSeek-V4 preview and publishes model cards and technical reports.

What's actually on offer: chat, API, and now an agent harness

DeepSeek is three products under one site. deepseek.com's chat interface is free. The API platform (platform.deepseek.com, docs at api-docs.deepseek.com) is pay-as-you-go against the same underlying models. The newest addition, announced 10 September 2026, is DeepSeek Harness — a downloadable desktop app (macOS/Windows) and open-source framework for building coding/research agents, positioned against tools like Claude Code, GitHub Copilot, and OpenCode. The API docs themselves point developers at Harness's quickstart guide and also document direct integrations with Claude Code and OpenCode for teams who just want to swap in DeepSeek as the backend model.

Current model: DeepSeek-V4.1-Flash

Released 10 September 2026, DeepSeek-V4.1-Flash is a 552B-parameter multimodal MoE model using a new 'Causal Encoder-Decoder' architecture: an 8B-parameter encoder pass on input tokens and a 16B-parameter decoder pass on output, rather than one dense activation path. DeepSeek says this cuts the KV cache footprint to roughly 1/4 of the prior V4-Flash and 1/8 of its SSD footprint, which matters for agentic workloads where repeated cache hits dominate cost. It natively handles images and text and supports up to 1M tokens of context. On DeepSeek's own published benchmark table, V4.1-Flash beats Claude Opus-5.0, GPT-5.6, Kimi K3 and GLM-5.3 on several agentic tasks (Terminal-Bench 2.1, DeepSWE v1.1, CyberGym, AutomationBench, Agent's Last Exam) but trails Opus-5.0 and GPT-5.6 on raw reasoning (HLE: 36.8 vs Opus's 56.3) and on the newer, harder Terminal-Bench 3.0/4.0 suites. These are DeepSeek's own reported numbers from its technical report, not independently reproduced.

API pricing: real numbers, peak/off-peak split, and a quiet model swap

Pricing is public on api-docs.deepseek.com and billed per 1M tokens with a peak/off-peak split (off-peak is half price; peak is 01:00-04:00 and 06:00-10:00 UTC weekdays, excluding Chinese holidays). For deepseek-flash: $0.003-$0.006 per 1M input tokens on a cache hit, $0.15-$0.3 on a cache miss, $0.6-$1.2 per 1M output tokens. For deepseek-v4-pro, the same tiers run roughly 7x higher ($0.022-$0.044 / $0.66-$1.32 / $1.98-$3.96). One thing worth knowing before you build against it: DeepSeek is deprecating V4-Pro. As of 14 September 2026, requests to deepseek-v4-pro are silently routed to V4.1-Flash and billed at the cheaper Flash rate, pending a future V4.1-Pro. The legacy model names deepseek-v4-flash and deepseek-v4-flash-vision-exp are likewise retired and now route to V4.1-Flash.

DeepSeek Harness: a serious agent-coding play, but still a preview

DeepSeek Harness is a separate open-source project (github.com/deepseek-ai/deepseek-harness, MIT license) built on the Cordis 'everything is a plugin' framework, where models, tools, skills, sandboxes, and the UI itself are all swappable plugins. It ships as a desktop app and can also run headless/as a web service from code. Features include an experimental multi-agent 'team' mode and an experimental auto-authorization review flow for tool calls. The repository is extremely active (pushed within the last three days as of this review) and has accumulated roughly 248,000 GitHub stars and 29,800 forks since its 13 August 2026 creation. DeepSeek itself labels it a developer preview on both the product page and the API docs, and the download page states plainly that 'the current version is still being iterated on, and some features and experience are still being refined' — so treat it as something to evaluate, not to deploy as a finished tool yet.

Licensing across the model lineup

DeepSeek's recent flagship releases are open-weight under a plain MIT license — confirmed directly on the DeepSeek-V4.1-Flash Hugging Face model card ('This repository and the model weights are licensed under the MIT License') and on the DeepSeek-V3 GitHub repo. That's unusually permissive for a frontier-scale model; there's no field-of-use or redistribution restriction to negotiate around, unlike many competing labs' 'open-weight' releases. DeepSeek Harness is separately MIT-licensed.

Who's behind it, and where it can't be used

DeepSeek (legal name Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd.) is owned and funded by High-Flyer, a Chinese quantitative hedge fund co-founded by Liang Wenfeng, who also serves as DeepSeek's CEO; the company was spun out of High-Flyer's AI research lab in July 2023. Independent reporting (via Wikipedia, citing Bloomberg and the Financial Times) describes a May 2026 Series A of $7 billion at a $52 billion post-money valuation, and later-stage talks in July 2026 targeting a $70 billion pre-money valuation alongside IPO preparations ; we found no primary filing or DeepSeek announcement that confirms these figures, so treat them as reported rather than confirmed. Two restrictions matter for buyers in regulated sectors: the US National Defense Authorization Act for Fiscal Year 2026 directs the Department of Defense and intelligence community to remove DeepSeek-developed or High-Flyer-owned AI from their systems (with narrow research/national-security carve-outs), and Australia's Department of Home Affairs has issued a government-wide directive to block and remove DeepSeek products from Australian government systems.

What we checked on DeepSeek

YesPartlyNo

What it costs to startYes

What the price counts

API use is metered by token volume, with separate cache, input, output, and time-of-day rates.

Model prices are listed in units of per 1M tokens.

api-docs.deepseek.com/quick_start/prici…Checked

How you build with itNo public code found

Open source

47 public repositories on GitHub, none of them the product's code

On the page: github.com/deepseek-ai repositories

github.com/deepseek-aiChecked by us,

Questions about DeepSeek

How much does DeepSeek cost?
DeepSeek costs $0.003 per 1m input tokens. From api-docs.deepseek.com/quick_start/pricing, checked 29 October 2026.
Does DeepSeek have an API or an MCP server?
DeepSeek has an API.
What have we checked on DeepSeek?
1 fact, under one heading: What it costs to start.
How do we check DeepSeek?
We hold 1 fact on DeepSeek, quoted from 1 page. Facts and price last checked 29 October 2026. Last change found 1 April 2026.

How we check every listing

Sources and changes

Last full check

DeepSeek price history

We have read the price once so far, on 29 October 2026: from $0.003 per 1m input tokens.

PricePlanFirst seenLast seenReadingsSource
$0.003 per 1m input tokens29 October 202629 October 20261Pricing page, read 29 October 2026

From our research on 29 October 2026. A date is when we read the page, not when the price changed.

Similar to DeepSeek

Is DeepSeek yours?

Free, and the page then carries the verified owner mark. You confirm it with an email at deepseek.com. Claim this listing

Report an error if a fact is wrong.