DeepSeek — The AI lab that reset price expectations: a free chat app, a cheap API, and open weights you can run locally.
What it is
DeepSeek is an AI research lab offering a free chat assistant, a low-cost API with 1M-token-context models, and open-weight model releases that anyone can download and run locally. Its API speaks both OpenAI and Anthropic formats, with off-peak pricing at half rate.
Editor's review
Long-form introduction by the BetterPicker editors · checked against the official site · Oct 28, 2026
DeepSeek is the model family that made cost the story. Cleo Abram's explainer of the moment it broke through (6.2 million views) and NetworkChuck's guide to the only way he recommends running it - locally (1.25 million views) - frame the two halves of its reputation, and Fireship's 2026 return video (1.23 million views) shows the industry still flinches at each release. The official stack today: a free chat app, and an API whose current models are deepseek-flash (V4.1-Flash) and deepseek-v4-pro, both with 1M-token context, thinking and non-thinking modes, and endpoints that mimic OpenAI and Anthropic so existing code plugs in. Pricing is the hook - off-peak rates are half of peak, cached input on Flash costs $0.003 per million tokens off-peak versus $0.30 for uncached peak input, and output runs $0.60 to $1.20 per million - with Search and Fetch tool calls billed separately at $0.75 per thousand. The other hook is openness: flagship weights have been published, the R1 release lives on GitHub under an MIT license, and reviewers demonstrate running it through Ollama or LM Studio on their own machines with no outside connections at all - a fallback no closed frontier lab can offer.
What it does well
The price-performance is real: off-peak halves, cache discounts and per-million billing make it one of the cheapest capable APIs, and reviewers keep picking it as the default for budget-heavy workloads.
Open weights change the calculus: published models mean your fallback is a local copy, which no closed frontier lab offers, and the community around quantized local runs is large and active.
Integration is painless: OpenAI- and Anthropic-compatible endpoints, JSON output, tool calls and a Responses-style API mean swapping it into existing pipelines is usually a base-URL change.
Who it's for
Developers watching per-token spend, tinkerers with the GPUs to run open weights, and anyone building on OpenAI-compatible tooling who wants a cheaper engine behind the same interface. Privacy-first users should read the next section before touching the hosted service.
Where it falls short
The hosted service answers to Chinese law: reviewers who otherwise like the models point out that data on the official app and API sits under PRC jurisdiction, which is exactly why the loudest local-run guides exist.
Capability gaps remain: vision input is on Flash but not the Pro model, and peak-hour pricing concentrates in UTC working hours, so the cheapest rates require scheduling around them - painless for batch jobs, a real constraint for interactive work.
The release cadence has been turbulent: model renames, retired versions and relaunches mean production users should pin versions, read the changelog before upgrading, and keep a regression check in their pipeline for the day behavior shifts.
Specs at a glance
Facts from the official site · not editorial opinion
| Latest version | deepseek-flash (DeepSeek-V4.1-Flash-1017) and deepseek-v4-pro (DeepSeek-V4-Pro-0813); open-weight R1 on GitHub |
|---|---|
| Price | Chat app free; API from $0.003 per 1M cached input tokens (Flash, off-peak) to $3.96 per 1M output (Pro, peak); off-peak is half rate |
| Platforms | Web and mobile chat apps; REST API (OpenAI and Anthropic compatible); open weights for local runs via Ollama, LM Studio and similar |
| Open source | Model weights published openly (R1 under MIT, about 92,000 GitHub stars); service code is not open |
| Core capability | 1M-token context, thinking and non-thinking modes, JSON output, tool calls, FIM completion, vision on Flash |
| Integrations | OpenAI and Anthropic API formats, Responses API, Search and Fetch tools ($0.75 per 1,000 requests), third-party local runtimes |
| Best for | Cost-sensitive developers and local-model tinkerers who want frontier-adjacent quality at commodity prices |
Frequently asked questions
▸Is DeepSeek free to use?
The chat assistant is free on web and mobile. The API is pay-as-you-go per million tokens, with off-peak rates at half price and large discounts for cached input.
▸Can I run DeepSeek models locally?
Yes. Flagship weights have been published openly - R1 under an MIT license - and guides routinely run them through Ollama or LM Studio on consumer hardware, fully offline.
▸Is my data safe with DeepSeek's hosted service?
The hosted app and API operate under Chinese jurisdiction, a concern reviewers raise repeatedly. If that matters, use the open weights locally instead of the cloud service.
▸How does DeepSeek's API pricing compare to OpenAI's?
It is dramatically cheaper, especially off-peak and with cache hits - Flash cached input is $0.003 per million tokens off-peak. Exact savings depend on your traffic pattern and peak-hour overlap.
Video reviews
2 review videos aggregated · praise and criticism included alike · click through to the original video
Channels that covered it
Channels are aggregated as sources only — we don’t rate creators
Related tools
Where to go next
External links open in a new tab; external content is independent of this site.
Link down? Every object page is re-checked monthly.