DeepSeek API Pricing in 2026: The Full Cost Breakdown

As of October 19, 2026, DeepSeek API pricing has peak and off-peak rates. deepseek-flash (V4.1-Flash) costs $0.15 per million input tokens and $0.60 per million output tokens off-peak, doubling to $0.30 and $1.20 at peak, while deepseek-v4-pro costs $0.66 input and $1.98 output off-peak, or $1.32 and $3.96 at peak (official pricing page). Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays; every other hour, including weekends, bills at the off-peak half rate, and cache hits cost $0.003 to $0.044 per million tokens.

Last checked: October 19, 2026, against DeepSeek's official Models & Pricing page and release notes. Prices changed twice since July (August 16 and September 10), so recheck before you budget.

DeepSeek API pricing is still among the lowest for near-frontier AI, but it is no longer the flat rate card most articles quote. On August 16, 2026 DeepSeek introduced peak and off-peak pricing and raised V4-Pro's rates; on September 10 it replaced V4-Flash with V4.1-Flash at lower prices. Against Western flagships DeepSeek still wins by roughly 3x to 25x per token, depending on the model and the hour, but small Western models now undercut its Flash line. The full picture is more nuanced, because sticker price is not the whole cost. Here is everything, verified against DeepSeek's official docs and each provider's own pricing page.

DeepSeek API pricing at a glance (October 2026)

All prices are per 1 million tokens, from the official pricing page:

Both models have a 1 million token context window and a 384K max output. deepseek-flash allows 2,500 concurrent requests and accepts images; deepseek-v4-pro allows 500 and does not support vision.

One inconsistency to know about before you budget on V4-Pro: DeepSeek's September 10 release note said that from September 14, deepseek-v4-pro requests would route to V4.1-Flash at Flash rates until a V4.1-Pro launches. The pricing page, checked October 19, still lists deepseek-v4-pro as DeepSeek-V4-Pro-0813 at the Pro rates above. I have not been able to confirm which one applies to a given call.

DeepSeek also now sells Search and Fetch APIs at $0.75 per 1,000 successful requests each. Calls made through server-side tools in DeepSeek Harness, Claude Code and Codex are free until billing starts at 00:00 Beijing time on November 1, 2026 (pricing page, October 17 release note).

The three details in this table people miss

The cache-hit discount is enormous. A deepseek-flash cache hit costs $0.003 per million input tokens off-peak, 50x less than a cache miss; on V4-Pro the gap is 30x ($0.022 versus $0.66). If your workload reuses long prompts (system prompts, retrieved documents, codebases), your effective input cost can approach zero.

Output costs 3 to 4x input. Flash output is 4x its input price and V4-Pro output is 3x. Workloads that generate long responses (reports, code, articles) cost more than the input price suggests.

Time of day now changes the bill. Every rate has a peak and an off-peak price, and off-peak is half of peak. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday, excluding Chinese public holidays, which is 9:00 to 12:00 and 14:00 to 18:00 in Beijing and overnight for US teams. Batch work you can schedule off-peak costs half as much.

What changed in DeepSeek's pricing in 2026

The V4 era replaced everything you may have read about DeepSeek pricing in 2025, and the rate card has moved again since July. Three concrete changes matter:

The old model names are gone

DeepSeek's V4 release note scheduled the legacy `deepseek-chat` and `deepseek-reasoner` names for full retirement after July 24, 2026, and the current docs list only `deepseek-flash` and `deepseek-v4-pro`. On September 10, V4.1-Flash replaced V4-Flash: the older `deepseek-v4-flash` and `deepseek-v4-flash-vision-exp` names still work, but they are served by DeepSeek-V4.1-Flash-1017 and billed at the Flash price.

Peak and off-peak pricing is back

DeepSeek's 2025 off-peak program, which offered 50% off V3 and 75% off R1 during a daily 16:30 to 00:30 UTC window, officially ended on September 5, 2025 (DeepSeek V3.1 release notes). A new version arrived with the V4-Pro general release: from 16:00 UTC on August 16, 2026, every rate has a peak and an off-peak price, with off-peak 50% lower (V4-Pro release note). The same change raised V4-Pro's output price from a flat $0.87 to $1.98 off-peak and $3.96 at peak, which ADTmag covered on August 20. Flash prices then came down on September 10 with V4.1-Flash.

Reasoning is now a mode, not a separate model

V4 models expose thinking and non-thinking modes under one price, instead of the separate `deepseek-reasoner` pricing of the R1 era. Budgeting is simpler, but thinking mode consumes more output tokens for the reasoning trace, so reasoning-heavy calls still cost more in practice.

DeepSeek API pricing history: how we got here

DeepSeek's pricing has moved fast, and old numbers still circulate widely. The verified timeline:

The pattern to plan around: DeepSeek changes prices and model lineups more aggressively than Western providers, in both directions, and with short migration windows. If you build on it, budget engineering time for model migrations, not just tokens.

How the caching discount actually works on your bill

DeepSeek prices cached input tokens separately, and the difference is large enough to change architecture decisions. Consider a support assistant that sends a 20,000-token system prompt and knowledge context with every request:

• Without caching, 1,000 requests a day send 20 million input tokens at $0.15 per million on deepseek-flash off-peak: $3.00 per day, or $6.00 if they all land at peak.
• With the context cached, those same tokens cost $0.003 per million off-peak: about $0.06 per day ($0.12 at peak).

That is a 50x difference on the same workload. The practical rule: structure prompts so the stable content (instructions, documents, schemas) comes first and stays byte-identical across calls, and put the variable content (the user's question) at the end. This applies to most providers with prompt caching, but DeepSeek's cache-hit price is the most aggressive of the major APIs. If you want to model your own workload's numbers, the LLM cost calculator does this math across providers.

DeepSeek vs every major API: the full comparison

Per 1 million tokens, standard (non-cached) rates from each provider's official pricing page as of October 19, 2026. DeepSeek rows show off-peak / peak:

Sources, all checked October 19, 2026: DeepSeek, OpenAI, Anthropic, Google, xAI, Mistral, Alibaba Cloud Model Studio. Grok 4 Fast and Llama 4 Scout, in earlier versions of this table, no longer appear in xAI's model list or Together's serverless pricing.

Where DeepSeek actually wins

Against flagships, DeepSeek still wins by a wide margin. Even at peak, V4-Pro's input ($1.32) is about 7.6x cheaper than GPT-6 Astra or Claude Fable 5.1 ($10) and 3x cheaper than Claude Opus 5.5 ($4); its peak output ($3.96) is about 12.6x cheaper than Astra or Fable ($50) and 5x cheaper than Opus 5.5 ($20). Off-peak, those gaps double. Against the $2-input mid-tier (Claude Sonnet 5.5, GPT-6.1 Sol, Grok 4.7, Qwen3.8 Max), V4-Pro is about 1.5x cheaper on input at peak and 3x off-peak.

Where the gap disappears

Against small Western models, DeepSeek is no longer the cheapest option. GPT-6 Luna and Claude Haiku 5.5 ($0.10 / $0.50) and Gemini 2.5 Flash-Lite ($0.10 / $0.40) all undercut deepseek-flash even at its off-peak rate ($0.15 / $0.60). DeepSeek's real edge is not "cheapest model available." It is near-frontier capability below mid-tier prices, a distinction worth keeping straight when you are choosing a model.

A note on promotional prices

Two rows in the table carry end dates. Google lists Gemini 4 Argon at $2 / $10 through November 30, 2026, rising to $4 / $20 on December 1, and Gemini 3.8 Flash at $0.75 / $3.75 through December 31, rising to $1.50 / $7.50 on January 1, 2027. If you are budgeting past those dates, use the higher rate. Anthropic now lists Claude Sonnet 5 at $2 / $10 as a standard rate, so the introductory-price caveat in earlier versions of this post no longer applies.

This is the same pattern covered in the broader enterprise AI pricing comparison: the market has split into premium flagships, aggressive mid-tier options, and near-free small models, with a 50x or larger spread between the extremes. Where a task lands on that spread is a routing decision, which is why multi-model setups beat single-model setups on cost almost by default.

Three worked cost scenarios

Token math makes the differences concrete. All numbers use the official rates above. DeepSeek figures show all-off-peak and all-peak; real bills land in between.

Scenario 1: internal support assistant

10M input tokens and 2M output tokens per month, with 70% of input hitting DeepSeek's cache.

Scenario 2: agentic workload on a stronger model

50M input tokens (30% cache hit) and 10M output tokens per month. Agent workloads skew input-heavy because agents re-read context across multi-step tasks.

Why include Flash in an agentic scenario: DeepSeek's September 10 release note says tests by multiple parties put V4.1-Flash ahead of V4-Pro on performance, cost, speed and total runtime. If that holds for your workload, the Flash line is the one to price.

Scenario 3: high-volume lightweight classification

10M input, 2M output, no caching.

Scenario 3 is the honest one: for trivial tasks, small Western models now beat DeepSeek outright, even before peak pricing kicks in. DeepSeek's savings show up when you need near-frontier capability, as in scenario 2. Budgeting per user instead of per token? See how much enterprise AI should cost per user.

Is DeepSeek actually good? The benchmark picture, honestly

Cheap only matters if the model performs. The evidence is genuinely mixed depending on which benchmark you trust:

Coding benchmarks: strong

DeepSeek reports 80.6% on SWE-bench Verified for V4-Pro, the highest open-weights score and roughly tied with Gemini 3.1 Pro (SWE-bench leaderboard aggregation).

Harder agentic benchmarks: behind the flagships

On Scale AI's contamination-resistant SWE-bench Pro, V4-Pro-Max scores 55.4% against Claude Opus 4.7 at 64.3% and GPT-5.5 at 58.6% (same source).

Neutral composite indices: mid-tier

On the Artificial Analysis Intelligence Index, V4-Pro scores in the low 40s to low 50s depending on effort mode (mirror data), below Claude Opus 4.8 and GPT-5.5 tier scores.

A fair summary: V4-Pro is a legitimate frontier-adjacent model at a fraction of flagship prices, excellent for coding and high-volume work, but the top Western flagships still win on the hardest reasoning and agentic tasks. Teams report the practical difference matters most for long-horizon agent runs, where small per-step quality gaps compound, a dynamic that shows up across agent orchestration platforms generally. DeepSeek's popularity is real either way: roughly 173 to 175 million cumulative app downloads and about 130 million active users by the end of 2025 (Business of Apps).

Self-hosting DeepSeek: is running it yourself actually cheaper?

Because DeepSeek's weights are open, self-hosting comes up in almost every pricing discussion. The honest answer: for most teams, no, and the reasons are worth walking through.

What it takes to run DeepSeek yourself

DeepSeek's larger models are dense enough to require multiple high-memory GPUs to serve at usable speed, not a single consumer card. That means renting or buying data-center-class hardware (H100s or H200s), setting up an inference server, and staffing someone who can keep it running, monitor for drift, and handle upgrades every time DeepSeek ships a new version, which per the pricing history above happens every few months.

GPU rental costs in 2026

H100 rental rates vary widely by provider. Lambda Labs lists on-demand H100s at $3.29 per hour on its official pricing page, with reserved multi-GPU commitments as low as $1.99 to $2 per hour per unit. Across the broader market, rates range from $1.49 to $6.98 per hour depending on the provider and commitment length, according to a 15-provider comparison. Running even a modest multi-GPU cluster around the clock for a month runs into thousands of dollars before a single production token is served, before accounting for the engineering time to set it up and keep it patched.

The real math: self-hosted vs API

Compare that fixed monthly GPU bill against Scenario 2 above: an agentic workload costing about $43 to $86 a month on DeepSeek's own API. Self-hosting only starts to pencil out at sustained, high-volume usage well beyond what most teams run, and even then it trades a small, predictable API bill for GPU capacity you have to provision, monitor, and pay for whether or not it is fully utilized.

When self-hosting actually makes sense

Self-hosting is worth it when you have steady, very high-volume traffic that would otherwise cost more than a GPU cluster running continuously, when data residency rules require keeping model weights and inference entirely inside your own infrastructure, or when you already run GPU infrastructure for other workloads and can add inference to existing spare capacity. Outside those cases, DeepSeek's first-party API pricing is difficult to beat with a self-hosted setup once GPU rental and engineering time are counted honestly.

The costs that are not on the pricing page

For an individual developer, DeepSeek's API is cheap and capable, full stop. For an enterprise, four things belong in the total cost calculation.

Your data is stored in China

DeepSeek's own privacy policy states: "To provide you with our services, we directly collect, process and store your Personal Data in People's Republic of China" (DeepSeek privacy policy). Prompts, uploaded files, and chat history are covered. For regulated industries, that is usually disqualifying on its own.

Your inputs can be used for training by default

DeepSeek's terms allow using inputs and outputs (de-identified) to improve the service, with an opt-out toggle in the product (DeepSeek terms of use). Enterprises need that opt-out verified, not assumed.

Government restrictions keep expanding

At least 17 US states plus federal agencies including the Pentagon, NASA, the US Navy, and the Department of Commerce have banned or blocked DeepSeek on government systems, and Italy, Australia, Taiwan, and South Korea have restricted government use (The Conference Board). South Korea's privacy regulator separately found DeepSeek had transferred user prompts to a Beijing-based third party without consent (reported via WitnessAI). If your customers include government or regulated buyers, using DeepSeek's hosted API becomes a sales-cycle question, not just an infrastructure one.

The compliant workaround costs more than the sticker price

Because the model weights are open, you can run DeepSeek US-hosted via Amazon Bedrock or Microsoft Azure, which thousands of Bedrock customers already do (usage statistics). But hosted-cloud rates are meaningfully higher than DeepSeek's first-party API, which erodes the price advantage that brought you here, and it still leaves you managing infrastructure rather than calling an API. This tradeoff between cost, capability, and compliance is the central theme in choosing AI tools for enterprises with secure data.

Does DeepSeek have an enterprise or team plan?

As of October 2026, DeepSeek's official pricing page lists only the pay-as-you-go per-token rates above, plus per-request Search and Fetch pricing; there is no published enterprise tier, volume discount schedule, or team billing plan. That is a meaningful difference from Western providers, which typically offer negotiated enterprise agreements with committed-use discounts, dedicated support, and contractual data terms. Buying DeepSeek at scale today means either running on the standard API rate with no negotiated pricing, or moving to a hosted-cloud version (Bedrock, Azure) where the enterprise contract sits with the cloud provider, not DeepSeek directly.

When the DeepSeek API is the right choice

DeepSeek fits well when:

• You are a developer or startup optimizing API spend, without regulatory constraints on data location
• The workload is high-volume and cost-dominated: classification, extraction, summarization, code generation at scale
• You can structure prompts to exploit the 30 to 50x cache-hit discount, and schedule batch work into off-peak hours
• You are building agent workflows where token volume makes flagship pricing painful, and per-step quality gaps are acceptable

It fits poorly when:

• You handle regulated, customer, or confidential data that cannot be stored in China
• You sell to government or enterprise buyers who ask vendors about their AI supply chain
• Your hardest tasks need the top of the benchmark table, where Western flagships still lead
• You need contractual enterprise guarantees (SOC 2 reports, DPAs, admin controls, volume pricing) that a raw budget API does not provide

Getting DeepSeek-level economics without the tradeoffs

The reason DeepSeek's pricing resonates is not DeepSeek itself. It is that most teams quietly overpay by routing every task, trivial or hard, to one premium flagship model. The 50x price spread between model tiers means the biggest AI cost lever is routing each task to the cheapest model that does the job well, which the 2026 LLM pricing analysis walks through, without the quality cliff of going budget-only.

That is the model behind Coworker: it routes each task to the right model automatically, so simple work runs on fast, cheap models and hard work runs on frontier models, with token cost 51x lower using OM2 plus model routing in Coworker's 100-task benchmark. Unlike a raw budget API, it is SOC 2 Type II certified and GDPR compliant with models hosted in the US, connects to 50+ enterprise tools like Salesforce, Slack, Jira, and Google Workspace, and carries organizational memory across tasks, which is the part that makes AI actually useful at work rather than another failed pilot. If your goal is running real AI agent use cases across your tools instead of managing API keys and GPU clusters, gateways like OpenRouter or a DIY LLM gateway solve routing but leave integration and memory to you. And if you are still mapping the landscape, start with the difference between agentic and generative AI or how teams build agents without code.

For a wider view, this AI token cost comparison lines up list prices, cost per task and one monthly workload for 13 models.

Frequently asked questions

How much does the DeepSeek API cost in 2026?

As of October 19, 2026, deepseek-flash (V4.1-Flash) costs $0.15 per million input tokens on a cache miss, $0.003 on a cache hit and $0.60 per million output tokens off-peak, doubling at peak. deepseek-v4-pro costs $0.66 input, $0.022 cached and $1.98 output off-peak, or $1.32, $0.044 and $3.96 at peak. Both have 1M context windows, per DeepSeek's official pricing page.

What happened to deepseek-chat and deepseek-reasoner?

DeepSeek's V4 release note scheduled both names for full retirement after July 24, 2026, and the current docs list only deepseek-flash and deepseek-v4-pro. The older deepseek-v4-flash name still works and is served by V4.1-Flash at the Flash price.

Does DeepSeek still offer off-peak discounts?

Yes. Since August 16, 2026, off-peak rates are half the peak rates. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday, excluding Chinese public holidays; all other hours, including weekends, are off-peak.

How does DeepSeek pricing compare to GPT-6 and Claude?

At peak rates, deepseek-v4-pro is about 7.6x cheaper than GPT-6 Astra and Claude Fable 5.1 on input and about 12.6x cheaper on output, and off-peak those gaps double. Against small models like GPT-6 Luna, Claude Haiku 5.5 or Gemini 2.5 Flash-Lite, deepseek-flash is now more expensive, so DeepSeek's advantage is strongest where you need near-frontier capability.

Is self-hosting DeepSeek cheaper than using the API?

Usually not. H100 GPU rental runs $1.49 to $6.98 per hour depending on the provider, and DeepSeek's larger models need multiple GPUs to serve well, so a self-hosted setup can cost thousands of dollars a month before accounting for engineering time. That is typically far more than the equivalent workload costs on DeepSeek's own API. Self-hosting only makes sense at sustained, very high-volume usage or when data residency rules require it.

Does DeepSeek offer enterprise or volume pricing?

No. As of October 2026, DeepSeek's official pricing page lists only standard per-token rates, plus per-request Search and Fetch pricing, with no published enterprise tier or committed-use discount, unlike most Western providers.

Is the DeepSeek API safe for enterprise use?

It depends on your data. DeepSeek's privacy policy states user data is stored in China, its terms allow training on inputs by default with an opt-out, and 17+ US states plus federal agencies have banned it on government systems. Enterprises that cannot accept those terms can run the open-weight models US-hosted on Amazon Bedrock or Azure at higher cost, or use a multi-model platform with enterprise compliance built in.

How do I get DeepSeek-level costs with enterprise compliance?

Route tasks across model tiers instead of sending everything to one flagship. Multi-model routing captures most of the savings that make DeepSeek attractive versus flagship-only usage, while keeping data with providers that offer SOC 2, US hosting, and enterprise terms. Coworker does this routing automatically across 50+ connected enterprise tools.

Related reading

• LLM cost calculator and 2026 pricing report
• Enterprise AI pricing comparison 2026
• The cheapest way to access Claude, GPT, and Gemini
• Multi-model AI vs single model
• How much should enterprise AI cost per user?
• Guru pricing
• Moveworks pricing
• Why enterprise AI fails