DeepSeek is the Chinese lab whose consumer chat app remains completely free — no Plus tier, no Pro subscription, no message caps beyond fair-use throttling. The model lineup turned over in 2026: V4 Flash and V4 Pro replaced V3 and R1, both with 1M-token context, and the legacy deepseek-chat and deepseek-reasoner API aliases were retired on July 24. The API also stopped being flat-rate — as of August 16, 2026 it bills peak and off-peak, with off-peak roughly half of peak.
Reviewed by EdGrows·Updated Aug 18, 2026Affiliate link — we may earn a commission
Type "DeepSeek pricing" into a search box and the honest answer is one sentence: the app is free.
No Plus plan. No Pro subscription. No message counter, no locked features, no premium models held back. You get V4 Flash and V4 Pro, built-in web search, unrestricted file uploads and saved history at zero cost.
The only limit is fair-use throttling — during busy periods you'll see server-busy warnings that slow things down temporarily.
In a market where every other frontier lab has converged on $20 a month, a genuinely free assistant running competitive reasoning models is an anomaly worth taking seriously. It's also the correct first stop for anyone about to pay for AI: try this, then decide what the subscription is actually buying you.
Two changes catch out anyone working from older material.
V4 replaced V3 and R1. The current models are V4 Flash and V4 Pro, both carrying a 1M-token context window with up to 384K maximum output. If you're reading a comparison that discusses R1 as DeepSeek's reasoning model, it's describing the previous generation.
The legacy API aliases are gone. deepseek-chat and deepseek-reasoner were retired on July 24, 2026. New integrations call deepseek-v4-flash and deepseek-v4-pro directly. A related trap: on the hosted API, deepseek-reasoner used to map to Flash's thinking mode rather than to Pro, which caused confusion even before it was removed.
On August 16, 2026, DeepSeek moved from a single published rate to peak and off-peak billing.
| Model | Off-peak (in/out per 1M) | Peak (in/out per 1M) |
|---|---|---|
| V4 Flash | ~$0.22 / $0.66 | ~$0.44 / $1.32 |
| V4 Pro | ~$0.66 / $1.98 | ~$1.32 / $3.96 |
Peak windows are 01:00–04:00 and 06:00–10:00 UTC; everything else is off-peak.
Two things follow. First, batch work that can be scheduled should run off-peak — it's a straight 50% saving for moving a cron job. Second, even at peak rates this remains among the cheapest frontier-adjacent APIs available, with V4 Flash output running an order of magnitude below comparable Western models.
The bigger lever is cache-hit pricing. Repeated prompt prefixes cost a fraction of fresh input — on Flash the difference runs from roughly $0.007 per million cached against $0.22 uncached. Structuring prompts so system instructions stay identical across calls is why production bills routinely come in far under the headline rate. DeepSeek describes caching as best-effort rather than guaranteed, so treat it as an optimisation rather than a promise.
Worth noting: the lab has signalled a future price increase without publishing replacement rates or a date. Budget on today's numbers and check before committing to volume.
Running DeepSeek alongside the paid assistants for everyday work, the thing that stands out is how little the free label costs you on technical tasks.
On maths, logic puzzles and structured coding problems it holds up against subscriptions costing $20 a month — the reasoning mode works through problems methodically and the visible chain of thought makes it easy to spot where an answer went wrong, which is genuinely useful when checking work rather than just accepting output. For debugging and explaining unfamiliar code it has been a reliable first stop.
The 1M-token context is the underappreciated part. Dropping in long documents or a substantial chunk of a codebase and asking questions across all of it works without the truncation dance that smaller windows force.
Where it feels less finished is everything around the model. The interface is plain, there are no integrations to speak of, and during busy hours the server-busy messages are frequent enough to interrupt a working session. Long open-ended creative writing is competent rather than distinctive — this is not where it wins.
The practical pattern that emerged: DeepSeek for technical reasoning and anything where a free tool is enough, paid assistants for work that touches client material or needs ecosystem integration. That split has less to do with quality than with jurisdiction, which is the next section.
DeepSeek is a Chinese company. Data goes to servers under Chinese jurisdiction.
For personal use — learning, coding practice, general questions, exploring ideas — most people reasonably treat that as a non-issue. For client work, regulated data, proprietary code or anything under an NDA, it's disqualifying no matter how good the model is or how attractive the price.
This is not a criticism of the product. It's the single fact that determines whether the free tier is a great deal or irrelevant to you, and it's worth deciding before you build habits around it.
The workaround exists: DeepSeek publishes open weights, so self-hosting removes the jurisdiction question entirely and gives fixed costs at volume. That requires serious GPU hardware, which is why it's an organisational answer rather than an individual one.
Confidential or client work. See above.
Ecosystem and integrations. No plugins, no connectors, no app marketplace. ChatGPT and Gemini are entire platforms; DeepSeek is a model with a text box.
Reliability during peak hours. Free means unprioritised. If you need guaranteed availability for time-sensitive work, pay someone.
Long-form creative writing. Competent, not distinctive. Claude remains the stronger writing partner.
Against ChatGPT: ChatGPT has the ecosystem, the integrations, the reliability and a Western data posture, for $19.99 a month. DeepSeek matches it closely on technical reasoning for nothing. The honest test: try DeepSeek first and see what you actually miss — for a lot of people the answer is integrations rather than intelligence.
Against Claude: Claude is better at writing, nuance and sustained long-form work, and its data handling suits professional use. DeepSeek competes on maths and code at zero cost. Different jobs, and running both is free on one side.
Against Gemini: Gemini bundles Workspace integration, video generation and storage at $19.99, or a capped tier at $4.99. DeepSeek offers none of that and costs nothing. If you're paying for integration, that's a rational purchase; if you're paying for raw reasoning, compare carefully.
Against Mistral: Mistral is the European answer to the same jurisdiction question DeepSeek raises — open weights, EU data residency, and paid tiers. For organisations that can't use Chinese infrastructure but want open models, that's the direct substitute.
Against Perplexity: Perplexity is a citation-first research engine rather than a general assistant. DeepSeek's built-in web search is functional but nowhere near as rigorous about sourcing.
| Route | Cost | Notes |
|---|---|---|
| Web and mobile app | Free | No paid tier, fair-use throttling only |
| API — V4 Flash | ~$0.22/$0.66 off-peak · ~$0.44/$1.32 peak | Per 1M input/output tokens |
| API — V4 Pro | ~$0.66/$1.98 off-peak · ~$1.32/$3.96 peak | Roughly 3x Flash |
| Cache-hit input | From ~$0.007/1M (Flash) | Best-effort, not guaranteed |
| Self-hosting | Hardware only | Open weights, removes jurisdiction issue |
Checked August 2026. DeepSeek moved from flat-rate to peak/off-peak API billing on August 16, 2026 — peak windows are 01:00–04:00 and 06:00–10:00 UTC. The V4 generation replaced V3 and R1, and the deepseek-chat and deepseek-reasoner aliases were retired July 24, 2026. The lab has signalled a future price increase without publishing rates or an effective date. The consumer app remains free. Verify on DeepSeek's official pricing page before budgeting production workloads.
Schedule batch work off-peak. A 50% saving for changing a cron schedule is the easiest optimisation available.
Engineer your prompts for cache hits. Consistent system prompts and stable prefixes can cut input costs by an order of magnitude. This matters more than which model you pick.
Decide the jurisdiction question first. Everything else on this page is irrelevant if your data can't go to China.
Try the free app before paying anyone. Whatever you conclude, you'll know what a $20 subscription is actually buying you.
DeepSeek remains the most generous offer in consumer AI, and the generosity is real rather than a trial. A free app with no paid tier, no message caps, a million-token context window and reasoning quality that holds its own against $20 subscriptions on maths, logic and code — that combination doesn't exist anywhere else, and it's the correct starting point for anyone deciding whether to pay for AI at all.
The complications are on the developer side and in the fine print. The V4 generation replaced V3 and R1, and the old API aliases stopped working in July. API billing moved to peak and off-peak rates on August 16, roughly doubling costs during two daily windows. A further price increase has been signalled without published rates. None of that touches the free app, but it means anything built on DeepSeek needs its assumptions rechecked.
The constraint that actually decides this, though, isn't pricing. Data goes to servers under Chinese jurisdiction, which makes DeepSeek excellent for personal and learning use and unusable for client, regulated or proprietary work. Self-hosting the open weights is the only route around it.
For individuals, students and anyone doing technical work on their own material, recommend without reservation — it costs nothing to find out. For professional work involving other people's data, the answer is no regardless of quality.
Note: DeepSeek does not have an affiliate program. AIVario earns no commission from this page, and the rating carries no commercial incentive.
Best for: Students and self-learners, technical reasoning and coding practice, cost-sensitive developers, anyone evaluating whether a paid AI subscription is worth it Not ideal for: Client or regulated work, confidential code and documents, users needing integrations and plugins, time-critical work requiring guaranteed availability Bottom line: The best free AI assistant available, with genuine frontier-adjacent reasoning and no upsell — provided your data can go to Chinese servers, which is the only question that matters here.
Yes. The web chat and mobile app cost nothing, with no Plus plan, no Pro subscription and no paywalled features. You get full access to V4 Flash and V4 Pro, built-in web search, file uploads without restrictions and saved chat history. The only limit is fair-use throttling — during peak hours you may see server-busy warnings that temporarily slow things down. There is no consumer tier to upsell you to, which remains unusual among frontier labs.
They were superseded by the V4 generation. The current models are V4 Flash and V4 Pro, both with a 1M-token context window and up to 384K maximum output. Importantly for developers, the legacy deepseek-chat and deepseek-reasoner API aliases were retired on July 24, 2026 — new integrations must call deepseek-v4-flash and deepseek-v4-pro directly. Any tutorial using the old names is describing endpoints that no longer route.
It depends on the hour, which is new. DeepSeek moved to peak and off-peak billing on August 16, 2026: V4 Flash runs roughly $0.44 input and $1.32 output per million tokens at peak, and about $0.22 and $0.66 off-peak. V4 Pro is roughly three times that. Peak windows are 01:00–04:00 and 06:00–10:00 UTC. Even at peak rates it remains among the cheapest frontier-adjacent APIs available.
The single biggest lever on your DeepSeek bill. Repeated prompt prefixes — system prompts, standing instructions — are cached, and cached input costs a fraction of fresh input. On V4 Flash the gap runs from around $0.007 per million for cache hits against $0.22 for misses, which is roughly a thirtyfold difference. Structuring prompts with consistent prefixes to maximise cache hits is why real bills often come in far below the headline rate.
Not for anything confidential, and this is the honest constraint rather than a quality issue. DeepSeek is a Chinese company and data goes to servers under Chinese jurisdiction. For personal use, learning, coding practice and general questions that's a non-issue for most people. For client work, regulated data, proprietary code or anything under an NDA, it's disqualifying regardless of how good the model is. Self-hosting the open weights is the workaround.
Closer than the price difference suggests, particularly on maths, logic and coding. DeepSeek's reasoning models handle structured problems competitively with paid frontier models, which is why the free tier gets so much attention. Where it trails is polish, ecosystem, integrations and consistency on open-ended creative work. The honest summary: comparable on the hard technical tasks, behind on everything surrounding them.
Yes, for the open-weight releases, which is why the lab matters beyond its own app. Self-hosting removes the jurisdiction problem entirely and gives you fixed costs at volume, at the price of needing serious GPU hardware. For organisations that want frontier-adjacent reasoning without sending data anywhere, this is the route — and it's the main reason DeepSeek's releases reshape the open-model landscape each time.
Efficiency plus strategy. DeepSeek's architecture is genuinely more compute-efficient than several Western equivalents, and the lab prices aggressively to build adoption and developer mindshare. The company has signalled future price increases without publishing rates or dates, and the August 16 move to peak billing is the first visible step. Budget on current published rates and re-check before committing to volume.
The best sourced-answer engine there is, and the browser is now free. Pro at $20 is the tier; Max at $200 is a different product.
The fastest way to see what the literature says about a claim — provided you read the meter as a signal, not a verdict.
Google's source-grounded research tool. The free tier limits how much you use, not what you can do.