DeepSeek Pricing 2026: Complete Guide to API Costs, Free Tier and Plans

DeepSeek pricing in 2026 spans four tiers: a free web chat at chat.deepseek.com, the flagship DeepSeek V4 Flash API at $0.22/$0.66 per million tokens off-peak, DeepSeek V4 Pro at $0.66/$1.98 per million tokens off-peak, and a free API grant of 5 million tokens for every new developer account. All V4 models ship with a native 1 million token context window at no extra charge, which puts V4 Flash 36x cheaper on input and 107x cheaper on output than a top-tier model like GPT-5.6 Sol or Claude Opus 5. Against the cheap tiers the gap is much narrower, and since OpenAI’s July price cut it is 1.4x on input.

This guide covers every part of DeepSeek pricing in 2026: the new V4 Flash and V4 Pro API rates, the legacy V3 chat and R1 reasoner costs, the free tier and the peak-pricing change DeepSeek has announced but not yet dated. We also cover OpenRouter and AWS Bedrock pricing, plus how DeepSeek stacks up against ChatGPT, Claude and Gemini. If you just want a single subscription that bundles DeepSeek with the other top models, we cover that too at the end. Our full DeepSeek vs Claude comparison puts the two side by side.

The Key Takeaways

• DeepSeek V4 Flash costs $0.22/M input off-peak and $0.44/M at peak (cache miss), with output at $0.66 and $1.32, in a 1M token context window.
• Prices rose on 16 August 2026. V4 Pro now costs $0.66/M input off-peak and $1.32/M at peak, with output at $1.98 and $3.96. The permanent 75% discount is gone.
• DeepSeek’s web chat is free for individual users with no Plus or Pro plan.
• The DeepSeek API gives 5 million free tokens to every new account.
• Cache hits cost a fraction of cache misses, so prompt caching can cut bills by 80% or more.
• The old 50 to 75% off-peak discount is gone. DeepSeek replaced it with 2x peak-hour pricing on 16 August 2026.

What is DeepSeek pricing?

DeepSeek pricing is pay-per-token: you are charged per million tokens of text the model reads (input) and writes (output). There are no monthly subscriptions on the API and no per-seat fees. Since 16:00 UTC on 16 August 2026 the rate also depends on when you call it. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, and every other hour is off-peak at exactly half the peak rate. Current rates start at $0.22 per million input tokens off-peak for DeepSeek V4 Flash.

DeepSeek deliberately undercuts every Western lab. Its closest rival on price is Alibaba’s Qwen, broken down in our Qwen pricing guide. As we covered in our DeepSeek V4 launch breakdown, the company’s stated goal is rock-bottom prices, backed by Huawei Ascend and Cambricon chips. That keeps DeepSeek pricing far below the flagship OpenAI and Anthropic tiers, though the cheap tiers at those labs have closed most of the gap on input tokens.

DeepSeek pricing 2026 (free chat, V4 Flash and V4 Pro at peak and off-peak rates)

DeepSeek pricing in 2026 covers four tiers: a free web chat, two API models (V4 Flash and V4 Pro), and a free 5-million-token grant for new developer accounts. All figures below come from the official DeepSeek pricing page and are in USD per 1,000,000 tokens.

DeepSeek web chat at $0 (free for individuals)

The DeepSeek web chat at chat.deepseek.com and the official mobile app are completely free for individual users. There is no Plus plan, no Pro tier, and no paywall on file uploads or long conversations. The only catch is fair-use throttling, so during peak hours you may see “Server Busy” warnings. Free chat sessions run on V4 Flash by default and let you toggle DeepThink to switch into the V4 Flash thinking mode.

DeepSeek V4 Flash pricing

DeepSeek V4 Flash is the default low-cost API model. It costs $0.22 per million input tokens off-peak on a cache miss and $0.44 at peak, with output at $0.66 and $1.32. Cache hits fall to $0.007 and $0.014 per million. V4 Flash supports both a non-thinking and a thinking mode under the deepseek-v4-flash ID and ships with a 1 million token context window plus up to 384K tokens of output. This is the model to default to for everything that does not require deep reasoning.

Since 31 July 2026 that ID serves DeepSeek-V4-Flash-0731, the official build that replaced the April preview. Artificial Analysis lifted V4 Flash from 40 to 50 on the Intelligence Index version it was running at the time. The price did move, though. The flat $0.14 / $0.28 rate this model launched on was retired on 16 August 2026, so any cost-per-task figure calculated before that date no longer holds.

DeepSeek V4 Pro pricing

DeepSeek V4 Pro is the flagship reasoning model. It costs $0.66 per million input tokens off-peak on a cache miss and $1.32 at peak, with cache hits at $0.022 and $0.044, and output at $1.98 and $3.96. V4 Pro shares the same 1M context and 384K max output as V4 Flash, and the model ID is deepseek-v4-pro. It is still far cheaper than any Western flagship, but the 75% discount DeepSeek once called permanent ended in August 2026.

DeepSeek free API grant at $0 (5 million tokens for 30 days)

Every new DeepSeek developer account starts with a 5 million token grant valid for 30 days, with no credit card required. That is enough for roughly 2,500 to 5,000 test calls depending on prompt length, worth about $8.40 at current V4 Flash rates. After the grant runs out, billing switches to the standard pay-per-token rates above and you can top up your balance with a small deposit.

DeepSeek per-million-token detail (cache miss vs cache hit)

The pre-V4 deepseek-chat and deepseek-reasoner model IDs are gone. They were retired after 24 July 2026, 15:59 UTC, and DeepSeek’s docs now list exactly two model IDs, deepseek-v4-flash and deepseek-v4-pro. If you still have those strings in a config file, the calls will fail. The V3 and R1 rates below are kept for historical reference only.

V4 Flash vs V4 Pro pricing

DeepSeek V4 Flash is the cheap, fast workhorse. V4 Pro is the deep-thinking model with a much larger active parameter count. V4 Pro costs roughly 3 times more per token than Flash, so default to Flash and escalate to Pro only when a task genuinely needs reasoning depth. Even at 3x, V4 Pro is the cheapest frontier reasoning model on the market.

How does DeepSeek pricing work?

DeepSeek pricing has three layers and you save money on each one.

1. Tokens. Every request is metered in input tokens (what you send) and output tokens (what the model generates). One million tokens is roughly 750,000 English words.
2. Cache hits. If part of your prompt has been seen recently, DeepSeek serves it from cache at a steep discount (about 98% off the cache-miss rate on V4 Flash). System prompts and shared context benefit the most.
3. Peak-hour surcharge, live. The V3 and R1 off-peak discount retired with those endpoints. DeepSeek replaced it on 16 August 2026 with peak and off-peak pricing where peak hours cost 2x the off-peak rate, across all billing items. Peak is defined as 9:00 to 12:00 and 14:00 to 18:00 Beijing time, which is 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday. Budget at the peak rate for anything that runs inside those windows.

You only pay for what you use. There is no minimum spend, no monthly fee, and no charge for context window length itself.

Is DeepSeek pricing free? Free tier and web chat

Yes, DeepSeek has free options. The consumer chat experience at chat.deepseek.com and in the official mobile app is completely free for individual users. There is no Plus plan, no Pro tier, and no paywall on file uploads or long conversations. The only catch is fair-use throttling, so during peak hours you may see “Server Busy” warnings.

For developers, the API includes a 5 million token grant for new sign-ups, valid for 30 days. That is enough for 2,500 to 5,000 test calls depending on prompt length, worth roughly $8.40 at current V4 Flash rates. No credit card is required, and after the grant runs out billing switches to the standard pay-per-token rates above.

DeepSeek pricing on OpenRouter, AWS Bedrock and Azure

If you do not want to use DeepSeek’s own API, several third-party providers host the same models. Pricing varies because each platform adds its own margin, infrastructure cost and routing.

OpenRouter also offers several free DeepSeek distilled models including the smaller R1 distills and DeepSeek V3 Base, with rate limits but no token charges. They are great for prototyping.

DeepSeek R1 vs V3 pricing

This is one of the most-asked questions about DeepSeek pricing. The short answer: R1 reasons better but costs roughly 5 times more per query. When it lands, we will map out DeepSeek R2 pricing as soon as official rates go public.

R1 generates many more output tokens per request because it produces a chain-of-thought before its final answer. Output is the dearer half of the bill, as our explainer on AI tokens sets out. That widens the price gap in practice. With V4, both Flash and Pro support thinking and non-thinking modes, so you can pick reasoning depth on a per-call basis. For more on the original reasoner, see our full breakdown of DeepSeek-R1 and how it beat OpenAI.

How DeepSeek pricing compares to GPT-5.6, Claude and Gemini

* Claude Sonnet 5 is $2/$10 on an introductory rate. It was due to go to $3/$15 on 1 September 2026, but Anthropic cancelled the increase on 10 August 2026. Note how much the ranking depends on tier: the frontier models are still 36x dearer on input, but OpenAI’s July 2026 cut dropped Luna to $0.20, which is close enough that raw token price is no longer the whole argument.

DeepSeek is not always the smartest model on a benchmark, but on price per useful token nothing else gets close. For high-volume workloads like content pipelines, summarization, classification and RAG over big corpora, DeepSeek is often the only model that makes the unit economics work. For the full side-by-side including monthly subscription tiers, see our complete AI pricing comparison. And for a head-to-head on capabilities, accuracy and where each model wins, read our DeepSeek vs ChatGPT comparison.

How to optimize DeepSeek API pricing costs

A few quick wins can cut a DeepSeek bill by 50 to 90 percent.

1. Pin your system prompt. Cache hits are 98% cheaper on V4 Flash. Keep the same opening context across requests.
2. Pick V4 Flash by default. Only escalate to V4 Pro when the task actually needs reasoning depth.
3. Cap your output. You pay per output token, so set max_tokens to the smallest value that still works.
4. Keep batch jobs outside Beijing office hours. The expensive windows are 9:00 to 12:00 and 14:00 to 18:00 Beijing time (UTC+8), which is 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday. Scheduling heavy jobs outside them halves the bill, and has done since 16 August 2026.
5. Use V4’s 1M context wisely. A long context window does not cost extra in DeepSeek’s pricing model, so you can stuff more retrieval results in instead of paying for repeat round-trips.
6. Try OpenRouter free distills first. Prototype on the free R1 distills, then switch the model ID when you’re ready to ship.

Want DeepSeek without dealing with API keys? Try Fello AI

Most readers do not actually want to write code against an API. They want to chat with the best model for the task without juggling five subscriptions, five web tabs and five different memory contexts.

Fello AI is the simplest way to do exactly that. One subscription gets you DeepSeek, ChatGPT, Claude, Gemini, Grok, Perplexity and several other top models in a single Mac, iPhone and iPad app for $9.99 per month. There is no per-token billing to track, no usage cap to budget around, and no separate account to manage for each provider. Everything lives in one chat window and one history.

Why Fello AI works well for DeepSeek users

DeepSeek is still one of the cheapest serious models on the API, but the official web chat is throttled and the desktop experience is bare-bones. Fello AI fixes both. You get DeepSeek inside a real Mac-native app with prompt history, file uploads, image analysis, voice input and the ability to switch models mid-conversation when DeepSeek is not the right tool for the next message. If your prompt would be better answered by Claude’s writing or GPT’s reasoning, swap models without losing context.

What does it actually save you

Add up the standard subscription cost of the major models: ChatGPT Plus at $20/month, Claude Pro at $20/month, Gemini Advanced at $20/month, Perplexity Pro at $20/month and Grok at $30/month. That is roughly $110 per month for individual access. Fello AI compresses the same lineup into $9.99 per month, with DeepSeek thrown in as part of the bundle. With 25,000+ five-star reviews across the App Store and Mac App Store, it is the most-loved AI bundle app in the Apple ecosystem.

If you would rather keep DeepSeek separate from the others, we also have a guide to a dedicated DeepSeek desktop client for Mac.

Conclusion

DeepSeek pricing in 2026 is still among the cheapest at the frontier, but it is no longer a flat rate. V4 Flash at $0.22/$0.66 off-peak, or $0.44/$1.32 at peak, still beats every Western lab on cost, and V4 Pro at $0.66/$1.98 off-peak keeps a top reasoning model under two dollars per million output tokens. What changed on 16 August 2026 is that your bill now depends on the clock. The web chat stays free, the API hands out 5 million free tokens, and prompt caching cuts repeat input costs by about 97% on top of that.

If you build software, go straight to the official DeepSeek pricing page and start with V4 Flash. If you just want to use DeepSeek alongside ChatGPT, Claude and Gemini in one place, Fello AI is the simplest route.

FAQ