DeepSeek Pricing (2026): Free Chat, API Under $1/M Tokens

DeepSeek chat costs $0. Its model API charges peak and off-peak rates: deepseek-flash costs $0.15 off-peak or $0.30 peak per million cache-miss input tokens, plus $0.60 or $1.20 for output; deepseek-v4-pro costs $0.66 or $1.32 for input, plus $1.98 or $3.96 for output. Default production work to Flash, schedule flexible work off-peak, and pay for Pro only when better accepted results justify its higher rate.

DeepSeek pricing at a glance

DeepSeek has one free consumer product, two paid API models, and separate Search and Fetch APIs. There is no monthly or annual self-serve API subscription: model-token charges and billable Search/Fetch requests come out of your account balance.

These numbers were verified against DeepSeek's live Models & Pricing page on October 17, 2026.

Since August 16, 2026, at 16:00 UTC, DeepSeek has used peak and off-peak API pricing. Peak hours are 01:00-04:00 and 06:00-10:00 UTC, Monday to Friday, excluding Chinese public holidays; every other hour is off-peak, including weekends and Chinese public holidays in full. On those weekdays, the off-peak windows are 00:00-01:00, 04:00-06:00, and 10:00-24:00 UTC. The rates below follow the September 10 Flash price adjustment; the October 17 model refresh left Flash pricing unchanged.

Move flexible batch work off-peak to halve the API token bill for the same model, token volume and cache-hit mix.

As of October 17, 2026, DeepSeek's Models & Pricing page lists Search: $0.75 per 1,000 successful requests; Fetch: $0.75 per 1,000 successful requests. Server-side calls to these APIs through DeepSeek Harness, Claude Code, or Codex are currently free, but billing starts at 00:00 Beijing time on November 1, 2026. Those agents will then incur tool-request fees alongside model tokens. The published Search/Fetch rates have no peak/off-peak split.

A token is the unit for model billing; Search and Fetch use successful requests as their billing unit. Input is everything sent to the model, including instructions, history, documents, and tool context. Output is everything the model returns. A cache hit is repeated input that DeepSeek can reuse at the lower rate; a cache miss is input it must process at the full rate.

The two API models share a 1M-token context window and a 384K maximum output. The pricing page lists no separate long-context surcharge inside those limits. Their operating limits differ: Flash lists 2,500 concurrent requests, while Pro lists 500, with limits applied per account.

DeepSeek chat is free, but the API is a separate bill

DeepSeek's consumer app costs nothing. Its official app announcement, read on October 17, 2026, still says it is 100% free, with no ads and no in-app purchases.

Free does not mean the API is free. The consumer app is an interface for a person; the API is the metered product software calls. DeepSeek's API page deducts usage from a granted or topped-up balance. A free chat account therefore does not turn programmatic traffic into free traffic.

The same product split catches buyers elsewhere: an app subscription and API billing can share a brand while remaining separate products, as the Perplexity pricing breakdown shows. For DeepSeek, the distinction is even cleaner because the app has no published subscription price while API model tokens and billable web-tool requests have separate charges.

Who never needs to pay

You never need to pay if your use is interactive, occasional, and satisfied by the official web or mobile app. That covers personal questions, ad hoc writing, document discussion, and students who do not need to connect DeepSeek to another product.

DeepSeek does not list a student-specific plan or discount on the pricing or app pages verified for this review. Ordinary consumer access already costs $0.

The boundary is automation. The moment a CRM, support workflow, internal tool, or customer-facing product needs to call DeepSeek without a person opening the app, budget for API tokens and any billable Search/Fetch calls. Also avoid treating free chat as production infrastructure: the free-app page does not promise unlimited capacity or publish a service level.

Flash is the production default; V4 Pro is an escalation lane

DeepSeek-V4.1-Flash-1017 should receive the first request in most production routes. It has the lower price and the higher listed concurrency limit. Both models now support the Responses API; Flash also supports vision, which Pro does not.

V4.1 Flash

Use deepseek-flash, which currently serves DeepSeek-V4.1-Flash-1017. Off-peak, Flash charges $0.15 per million cache-miss input tokens, $0.003 for cache-hit input, and $0.60 for output. At peak hours those rates are $0.30, $0.006, and $1.20. It supports thinking and non-thinking modes, JSON output, tool calls, DeepSeek's Anthropic-compatible API, and the Responses API. The live page lists a 1M context window, 384K maximum output, and 2,500 concurrent requests.

The old deepseek-v4-flash and deepseek-v4-flash-vision-exp names remain accepted aliases, but their retired models have been replaced by DeepSeek-V4.1-Flash-1017 and their requests are billed at the current Flash rate.

That combination makes Flash the sensible route for classification, extraction, summarization, first-draft generation, routine code changes, and bounded agent steps. These jobs have clear acceptance checks, so a failed output can be detected and escalated without letting a cheap mistake reach a customer.

Skip Flash as the only route when failure is expensive and hard to detect. A legal, financial, or customer-impacting recommendation should not choose its model from token price alone. The low rate buys room to evaluate; it does not replace evaluation.

V4 Pro

Use deepseek-v4-pro, which currently serves DeepSeek-V4-Pro-0813. Off-peak, Pro charges $0.66 per million cache-miss input tokens, $0.022 for cache-hit input, and $1.98 for output. At peak hours those rates are $1.32, $0.044, and $3.96. It has the same 1M context window and 384K maximum output, plus the same thinking modes, JSON output, tool calls, and Anthropic-compatible format. Its listed concurrency limit is 500.

Pro is the hard-tail route: the ambiguous request, the complex plan, the difficult coding task, or the second attempt after Flash fails a deterministic check. It is not the default merely because its name says Pro.

The current pricing table lists Responses API support for Pro as well as Flash. That removes the old endpoint limitation. Pro still lacks Flash's vision support, so an image-dependent workflow cannot treat the two as interchangeable.

DeepSeek's token bill is tiny; cost per accepted result still decides

The raw model-token expense is small enough that one avoidable human review can outweigh thousands of model calls. The useful formula is not cost per request. It is cost per accepted result.

For model-token cost on a request with no cache hits, use the rates for its pricing period:

cost = input tokens / 1,000,000 × input rate + output tokens / 1,000,000 × output rate

At 1,000 tokens, Flash costs $0.00015 for cache-miss input and $0.00060 for output off-peak, or $0.00030 and $0.00120 at peak. Pro costs $0.00066 for cache-miss input and $0.00198 for output off-peak, or $0.00132 and $0.00396 at peak.

Consider a support-drafting workflow with 2,000 input tokens and 500 output tokens per reply. At 1,000 replies, Flash costs $0.60 off-peak or $1.20 peak, and Pro costs $2.31 or $4.62, with no cache hits. The Pro premium is $1.71 off-peak or $3.42 peak per thousand replies.

Now add labor. Five minutes of review at an illustrative $60 hourly cost is $5. One extra review event costs more than the Pro premium for that entire batch in either period. The right question is therefore not, "Which model has the lower line item?" It is, "Which route clears the acceptance check with the least model spend plus review time?"

What a monthly DeepSeek bill looks like

DeepSeek has no seat-month to compare with a software subscription. Monthly cost is the traffic shape: how much context goes in, how much text comes out, how often prefixes hit cache, which hours the work runs, how many attempts an accepted result takes, and how many billable Search/Fetch calls it makes. The model-token examples below assume all usage falls in one time band; a mixed month needs each band's usage priced separately.

Once the agent-tool billing above starts, add successful Search requests / 1,000 × $0.75 + successful Fetch requests / 1,000 × $0.75 to the model-token total. For example, one successful Search and one successful Fetch for each of 1,000 agent tasks adds $1.50. DeepSeek's billing guide requires HTTP 200 with non-empty Search results or Fetch page content. It excludes failed calls, empty Search results, and timeouts; changing the number of Search results does not change the request price. Avoid redundant searches and repeated fetches when an existing result is sufficient. The following estimates cover model tokens only.

For a solo product prototype using 10M cache-miss input tokens and 2M output tokens in a month, Flash costs $2.70 off-peak or $5.40 peak, and Pro costs $10.56 or $21.12. The difference is $7.86 or $15.72. The model that produces a valid result with less debugging can still matter more than that gap.

For a support operation drafting 100,000 replies at 2,000 input and 500 output tokens each, the uncached bill is $60.00 off-peak or $120.00 peak on Flash, versus $231.00 or $462.00 on Pro. At an illustrative 80% input cache-hit rate, those totals fall to $36.48 or $72.96 on Flash and $128.92 or $257.84 on Pro. At off-peak rates, one hour of human review still exceeds the cached monthly Flash bill under the $60-per-hour assumption.

For 1,000 document-analysis jobs using 50,000 input and 2,000 output tokens each, the uncached bill is $8.70 off-peak or $17.40 peak on Flash, versus $36.96 or $73.92 on Pro. With 80% of input hitting cache, it becomes $2.82 or $5.64 on Flash and $11.44 or $22.88 on Pro. Long input does not automatically mean a large DeepSeek bill; repeated long input can make the unit economics unusually small.

These examples also show why a fixed monthly estimate copied from another company is useless. A support reply is output-heavy relative to its prompt. A document job is input-heavy. An agent can multiply both through retries and tool loops. Save the token mix and pricing period beside the dollar figure so finance can tell whether traffic, quality, scheduling, or a vendor rate caused the change.

The Flash-to-Pro break-even

Pro's uncached input rate is 4.4 times Flash, while its output rate is 3.3 times Flash. There is no single price multiple for every workload. With no cache hits and the same pricing period, the blended premium falls between those two ratios.

For the support example's 2,000-input and 500-output token mix, Pro costs 3.85x Flash in either period. Under that fixed mix, Pro must reduce billed attempts or token volume by 74.03% to break even on model spend alone. Change the input/output mix, cache-hit rate, or execution period and recalculate.

If Pro finishes a task in one attempt, Flash can average up to 3.85 comparable attempts before its token bill becomes higher in this example. Three Flash attempts cost less than one Pro attempt; four cost more. Human review can move that boundary sooner.

This routing rule also prevents a common procurement mistake. A team sees low input rates, sends every task to Pro, and calls the total cheap. The absolute bill may still be small, but an unmeasured premium across millions of calls is not a strategy. Flash should own the bounded base load; Pro should own the failures that justify it.

Context caching changes the bill only when prefixes truly repeat

DeepSeek's cache-hit rate is valuable, but it is not a discount you can assume in a spreadsheet. The context-caching guide says caching is enabled by default, works on complete persisted prefix units, and operates on a best-effort basis.

A stable system prompt followed by a stable reference document is a strong cache candidate. A later request can reuse that beginning and append a new question. Reusing a phrase from the middle of a different prompt does not qualify. The complete prefix must match a unit DeepSeek has persisted.

The cache also has a time dimension. Construction takes seconds, and unused entries are normally cleared within hours to days. A weekly report may miss a cache that a burst of near-identical requests would hit. DeepSeek does not guarantee a 100% hit rate.

At an illustrative 80% input cache-hit rate, the blended input cost per million tokens falls to $0.0324 off-peak or $0.0648 peak on Flash, and $0.1496 or $0.2992 on Pro. In the 1,000-reply support example, total cost falls from $0.60 to $0.3648 off-peak on Flash and from $2.31 to $1.2892 on Pro. Peak totals are $0.7296 and $2.5784 respectively.

That saves 39.20% of Flash's model-token bill and 44.19% of Pro's in this output-bearing workload. Output tokens do not receive the input cache discount, so an 80% input hit rate cannot be treated as an 80% total-bill reduction. Input caching also does not discount separate Search/Fetch request fees.

1. 1

   Keep the prefix stable

   Put durable instructions and repeated reference material first. Append the changing user request after that stable prefix instead of rebuilding the prompt in a different order.

2. 2

   Read the usage fields

   Track prompt_cache_hit_tokens and prompt_cache_miss_tokens from API responses. A projected hit rate is not a billing fact until these fields show it.

3. 3

   Price the observed mix

   Split measured input into hit and miss tokens, add output tokens, and calculate cost per accepted result using each pricing period's rates. Reprice when prompt structure or traffic cadence changes.

DeepSeek is cheaper than the API shortlist, before quality adjustment

DeepSeek wins the raw-token comparison against these three models. It does not automatically win the business outcome because these models differ in capability, tools, latency, and failure rate.

Use one normalized workload to make the sticker prices comparable: 1M total input tokens plus 250K output tokens, with no cache discount. That workload costs $0.30 off-peak or $0.60 peak on Flash, and $1.155 or $2.31 on Pro.

OpenAI GPT-5.6 Sol currently charges $4 input and $20 output per million tokens at standard short-context rates, with promotional pricing available at least through November 21, 2026. The normalized workload costs $9.00 when each request stays at or below OpenAI's long-context threshold, or 7.79x off-peak Pro and 3.90x peak Pro. Prompts above 272K input tokens trigger 2x input and 1.5x output pricing for the full request, and cache writes cost 1.25x uncached input.

Claude Sonnet 5 charges $2 input and $10 output per million tokens. The same workload costs $4.50, or 3.90x off-peak Pro and 1.95x peak Pro. Anthropic includes the 1M context window at standard pricing and offers a 50% Batch discount, so latency-tolerant work can narrow the gap.

Gemini 3.1 Pro Preview charges $2 input and $12 output per million tokens when a prompt stays at or below 200K. The normalized aggregate workload costs $5.00 if every request stays under that boundary, or 4.33x off-peak Pro and 2.16x peak Pro. Price the same aggregate workload using prompts above 200K and Google's rates rise to $4 input and $18 output, producing an $8.50 bill, or 7.36x off-peak Pro and 3.68x peak Pro.

The price spread is large, but the dollar spread is smaller than the multiples make it sound. Moving the normalized workload from Pro to GPT-5.6 Sol adds $7.845 against off-peak Pro or $6.69 against peak Pro. One human correction can erase that saving, while a reliable high-volume route can compound it.

Use the DeepSeek vs ChatGPT comparison when model quality, product features, and data handling matter more than the token table. Use the cheapest AI API shortlist when the job can accept smaller budget models and gateways. A pricing page can normalize rates; only your evaluation can normalize outcomes.

The hidden budget risk is repricing, not an annual lock

DeepSeek's peak/off-peak schedule makes execution time a budget input. Its live pricing page also reserves the right to adjust prices and recommends topping up for actual usage. It no longer carries the old warning of an undated, significant upcoming increase.

That changes the procurement decision. Store the pricing-page verification date beside every budget model, make the rate and execution period editable inputs, and define the price change that would trigger a reroute.

There is no monthly or annual API trap

DeepSeek lists usage-based deduction for model tokens and successful Search/Fetch requests, not a monthly seat or annual self-serve contract. Topped-up balance does not expire, according to its current Help Center. The API uses granted balance first when both granted and topped-up balances exist.

Unused balance is refundable through Billing > Refunds for online payments, according to DeepSeek's current Help Center; corporate bank-transfer refunds require a ticket. Its Open Platform terms qualify that promise: requests are reviewed, approved refunds cover the remaining unspent amount less necessary costs, and partial refunds or refunds of spent usage are not supported. That makes a modest top-up easier to manage than a large prepaid balance.

When balance is insufficient, the API returns HTTP 402 and tells you to top up. The operational risk is interruption. Set a balance alert before production traffic finds that boundary for you.

Integration and concurrency can cost more than tokens

Flash lists 2,500 concurrent requests and Pro lists 500. Both now support the Responses API, while only Flash supports vision. A buyer who ignores those differences may pay in queueing, fallback logic, or integration work even when the token bill looks modest.

The cheapest safe architecture keeps routing reversible. Use one internal interface, record the chosen model, token usage, pricing period, and billable Search/Fetch counts, and preserve a tested fallback. Editable rates turn a pricing change into a budget adjustment rather than an application rewrite.

DeepSeek's price history split Flash from Pro

DeepSeek's current tiers do not make every line item uniformly cheaper than its historic V3 schedule. Flash remains the volume-pricing route, but peak hours and Pro's higher rates change the comparison.

DeepSeek's V3 pricing announcement listed $0.07 cache-hit input, $0.27 cache-miss input, and $1.10 output per million tokens. Against those historic rates, today's off-peak Flash is 95.71% cheaper on cache-hit input, 44.44% cheaper on cache-miss input, and 45.45% cheaper on output. At peak hours, Flash cache hits are still 91.43% cheaper, but cache-miss input is 11.11% higher and output is 9.09% higher.

Pro moves differently. Off-peak cache-hit input is 68.57% cheaper than historic V3, but its $0.66 cache-miss input is 144.44% higher and $1.98 output is 80.00% higher. Peak Pro cache hits are 37.14% cheaper, while cache-miss input is 388.89% higher and output is 260.00% higher.

That history explains the current decision rule. Flash is the volume-pricing play. Pro asks the workload to earn a higher rate through better outcomes. Scheduling and future repricing can change both, so the verification date and time band belong next to the arithmetic.

The Monday move: measure one route before moving the budget

On Monday, take 100 sanitized requests from one bounded workflow and run a one-week routing test. Do not start with an enterprise-wide migration.

1. 1

   Define acceptance first

   Choose a result a reviewer can score consistently: valid schema, correct classification, passing tests, or an approved support draft. Remove customer, personal, and confidential data from the evaluation set.

2. 2

   Send the base load to Flash

   Run all 100 requests through deepseek-flash. Record input, output, cache-hit, cache-miss, retry, latency, acceptance, and peak/off-peak data for each request. Also record successful Search and Fetch calls so the budget includes their fees when billing starts.

3. 3

   Escalate only failures

   Send failed or ambiguous cases to deepseek-v4-pro. Measure whether Pro clears the quality floor and reduces failed work enough to earn its workload-specific premium. For the uncached support mix, that premium is 3.85x before human review.

4. 4

   Set two budget triggers

   Keep Flash as the default only while it clears the quality floor. Keep DeepSeek as the provider only while the live normalized cost stays below your tested fallback after any price change.

The output of this test is one decision-ready number: cost per accepted result. It combines model tokens, billable web-tool calls, and review time into a workflow budget finance can revisit without rebuilding the analysis.

Frequently asked questions

Is DeepSeek paid or free?

DeepSeek's official consumer app is free, with no ads or in-app purchases. API access is separate: model tokens use peak or off-peak rates, and Search/Fetch have per-successful-request fees. Calls through DeepSeek Harness, Claude Code, or Codex start incurring those tool fees on November 1, 2026, at 00:00 Beijing time.

How much does DeepSeek V4 Pro cost?

deepseek-v4-pro costs $0.022 per million cache-hit input tokens, $0.66 per million cache-miss input tokens, and $1.98 per million output tokens off-peak. Peak rates are $0.044, $1.32, and $3.96 respectively. It serves DeepSeek-V4-Pro-0813, with a 1M context window, 384K maximum output, and a listed concurrency limit of 500.

How much is DeepSeek per month?

Consumer chat costs $0 per month. The API has no fixed monthly price: split monthly cache-hit input, cache-miss input, and output tokens by peak/off-peak period, then multiply each total by its per-million rate. Add billable Search/Fetch usage at $0.75 per 1,000 successful requests for each API; the named agent integrations start paying those tool fees on November 1, 2026.

Does DeepSeek offer a student discount?

DeepSeek does not list a student-specific plan or discount on the official pricing or app pages verified on October 17, 2026. Students using the consumer web or mobile app already pay $0.

Can I get a refund from DeepSeek?

Unused API balance can be refunded. For online payments, use Billing > Refunds; corporate bank-transfer payments require a support ticket. Topped-up balance does not expire. Refunds are subject to review and necessary costs under the Open Platform terms; partial refunds and refunds of spent usage are not supported.

Did DeepSeek pricing change in 2026?

Yes. Peak/off-peak pricing took effect on August 16, 2026, at 16:00 UTC, and V4.1 Flash brought another pricing adjustment in September. Current peak hours are 01:00-04:00 and 06:00-10:00 UTC, Monday to Friday, excluding Chinese public holidays. All other hours are off-peak, at half the peak token rates. The October 17 Flash model refresh left those rates unchanged. Search/Fetch billing through DeepSeek Harness, Claude Code, or Codex begins November 1, 2026, at 00:00 Beijing time, at $0.75 per 1,000 successful requests for each API.

Should I choose V4 Flash or V4 Pro?

Choose the current Flash model, deepseek-flash, for the bounded base load and Pro for cases where it demonstrably improves accepted results. Pro's uncached premium is 4.4x on input and 3.3x on output in the same pricing period. At the support example's fixed token mix it is 3.85x, requiring 74.03% less comparable billed work to break even on model spend.

Get the AI Tools Map for Business Owners

The AI Tools Map for Business Owners turns model prices into a practical adoption stack, with the quality floor, routing role, and cost trigger for each tool. Subscribe to get the next edition free.