# DeepSeek API Pricing 2026: Cost per 1M Tokens By Desmond Achebe · 2026-10-15 · Source: https://www.activepieces.com/blog/deepseek-api-pricing-2026-cost-per-1m-tokens --- **No single number prices a run**, because DeepSeek bills each API model on **three separate per-million-token rates** (cache-hit input, cache-miss input and output), so anyone tallying costs from a workflow log, say one exported from [Activepieces](https://www.activepieces.com), has to split the tokens into those three buckets first. ![A printed workflow log sheet on a desk, with three separate stacks of coins beside it for cache-hit input, cache-miss input…](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/fe93ecdf-dc33-4d95-a861-90547b98e40b/deepseek-api-pricing-2026-cost-per-1m-tokens-ill-044827fd.webp) DeepSeek API pricing is the usage-based charge for calling DeepSeek's models, set in US dollars per million tokens and split into cache-hit input, cache-miss input and output rates, with reasoning tokens billed as output. Add three terms and divide by 1,000,000, and you have the cost of one run ([Api-docs](https://api-docs.deepseek.com/quick_start/pricing)). The terms are cached input tokens × hit rate, uncached input tokens × miss rate and output tokens × output rate. Reasoning tokens count as output, so a model that thinks at length bills more than the same answer without the thinking. With pay-as-you-go you hold a balance and each call deducts the tokens it used. There's no seat fee and no minimum term. ## DeepSeek API pricing per 1M tokens, checked today ### The pricing table for every model Right now, two API models are on sale. Start with what each one is: deepseek-flash is DeepSeek-V4.1-Flash, accepts vision input and is positioned as fast and cheap, while deepseek-v4-pro is DeepSeek-V4-Pro-0813, takes no vision input and is positioned as higher capability. Rates are dollars per million tokens, from the [DeepSeek pricing page](https://api-docs.deepseek.com/quick_start/pricing). **Peak figures are double the off-peak ones**. | Model | Period | Cache-hit input | Cache-miss input | Output | |---|---|---|---|---| | deepseek-flash | Off-peak | $0.003 | $0.15 | $0.60 | | deepseek-flash | Peak | $0.006 | $0.30 | $1.20 | | deepseek-v4-pro | Off-peak | not confirmed | $0.66 | $1.98 | | deepseek-v4-pro | Peak | not confirmed | $1.32 | $3.96 | ![A printed rate table on paper with five columns and rows grouped in pairs for deepseek-flash and deepseek-v4-pro, one pair…](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/35b09b6e-3973-4347-ab2c-0c22ca441588/deepseek-api-pricing-2026-cost-per-1m-tokens-ill-a161b7c6.webp) _Prices and plan limits checked against [docs.claude.com](https://docs.claude.com/en/docs/about-claude/models/overview) and [claude.com](https://claude.com/pricing) and [openai.com](https://openai.com/chatgpt/pricing) and [gemini.google](https://gemini.google/subscriptions) on October 14, 2026._ The Pro figures of 0.66 and 1.98 in the vendor comparison further down are this table's off-peak cache-miss and output rates. That comparison has a single input column, and the uncached rate is the one that fits it. I could not confirm Pro's cache-hit rate, so read it off the page before you model caching on Pro. If a step reads screenshots or scanned attachments, only the Flash model can take it, and whatever its rate, you can't use the Pro model. ### Web chat and app versus the API You can use DeepSeek's web chat and app for free, and they are a separate product from the API. Using them consumes no API balance, and API spend buys nothing there. ### How the check was done and when it goes stale I confirmed the model names and versions above against DeepSeek's own model listing on 14 October 2026. I wrote this section on 15 October 2026. If a new version suffix appears, or the model carrying vision changes, the table needs rechecking before you trust any estimate built on it. ## Leverage cache hits for lower DeepSeek rates **Price input at your blended cache rate, not the miss rate.** On V4-Flash, a 90% hit rate takes input from $0.15 to $0.0177 per million tokens, which is the number to put in your estimate once you recheck the table. ![DeepSeek Flash input price at different cache hit rates](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/a42d993a-204c-4a05-94ef-1770837a917b/deepseek-api-pricing-2026-cost-per-1m-tokens-sta-3a3e30b5.svg "Source: DeepSeek API Docs") ### How context caching is billed DeepSeek charges each input token one of two ways. A cache miss is full price, $0.15 per million off-peak, according to DeepSeek API Docs, meaning any uncached query costs you this amount. Where the start of your prompt matches one DeepSeek has recently seen, that's a cache hit, and it costs $0.003 per million, according to the DeepSeek API Docs. A hit is therefore 50 times cheaper than a miss, so your prompt design sets your input bill more than the headline rate does. Peak pricing doubles both figures, to $0.30 and $0.006, so peak-hour requests are twice as expensive. ### Effective input price at three hit rates Weight the miss price by misses and the hit price by hits, and you have the blended rate. The chart shows it at 0%, 50% and 90% hits, off-peak and peak; the point to take is that the drop is steep and not proportional to the hit rate. At 0% hits you pay the full $0.15, so a cold prompt gets no discount. At 50% the rate is $0.0765, which roughly halves the bill. At 90% it is $0.0177, about an eighth of the rate card. At peak, each of these doubles, so a 90% hit rate (about $0.035) still beats a 0% hit rate off-peak by a wide margin. Take 1,000 runs with a 20,000-token prompt, or 20 million input tokens. That costs $3.00 at 0% hits and $0.354 at 90%, so the same workload differs by about $2.65 depending on layout alone. Don't count on those hit rates. They depend on how often you call, how long the cache lives and how stable your prompt is, so measure your own before trusting the 90% row. Because caching matches from the start of the prompt, order decides what gets reused: 1. Put fixed material first: system instructions, schemas, examples, reference documents. 2. Put anything that changes per run last: the email body, the ticket text, the user's question. 3. Keep the fixed block byte-identical. A timestamp or request ID near the top breaks the match for everything after it. 4. Append to a conversation rather than rewriting earlier turns. ![A prompt document on a desk in two sections: a thick top block of system instructions, schemas, examples and reference…](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/83afc57e-c2bf-4cf0-838c-c8bc9a2278ef/deepseek-api-pricing-2026-cost-per-1m-tokens-ill-1e2d8db6.webp) Check the usage figures the API returns on a few real calls, and compute your own hit rate from them. ## What 1,000 automation runs cost per model Once you have your own usage figures, the arithmetic is short. 1,000 runs of a two-step flow cost about $0.43 on DeepSeek V4.1-Flash and about $0.09 on Qwen 3.7 Flash, so the cheaper model is **roughly five times less**. ### Token assumptions for each step I assume a summarised email is 1,500 tokens in and 200 out, and a classified ticket is 500 in and 20 out. Swap in your own averages, because long email threads will move the input side far more than the output side. Quoted as input–output dollars per million tokens, the rates come from [GetAPIPulse](https://www.getapipulse.com/top-10-cheapest-llm-apis.html). DeepSeek V4.1-Flash is $0.15–$0.60, the highest in this budget group. Qwen 3.7 Flash is $0.03–$0.13, the lowest, so it's the comparison here. ![Activepieces pricing page with four subscription tiers showing features and monthly costs](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/5780ea9c-6ecc-45f7-99f2-8707dfb01930/vercel-pricing-in-2026-how-the-costs-scale-scree-22533578.webp) In that listing, the next cheapest was GPT-5 nano at $0.05–$0.40, an older model; OpenAI's current efficient option is GPT-6 Luna. [Gemini](https://gemini.google/subscriptions) Flash-Lite at $0.10–$0.40 and [Claude](https://claude.com/pricing) Haiku 5.5 at $0.10–$0.50 both sit between the two extremes, so any of them undercuts DeepSeek on list price. ### Cost per 1,000 runs, side by side Look at the table and the gap is proportional, not dramatic: the cheaper model costs about a fifth at every volume, so the saving only matters once volume grows. | Step | Tokens in / out | DeepSeek, 1,000 runs | Qwen 3.7 Flash, 1,000 runs | 10k runs a month (DeepSeek / Qwen) | 100k runs a month (DeepSeek / Qwen) | |---|---|---|---|---|---| | Summarise email | 1,500 / 200 | $0.345 | $0.071 | $3.45 / $0.71 | $34.50 / $7.10 | | Classify ticket | 500 / 20 | $0.087 | $0.018 | $0.87 / $0.18 | $8.70 / $1.76 | | Total | 2,000 / 220 | $0.432 | $0.089 | $4.32 / $0.89 | $43.20 / $8.86 | At 10,000 runs a month the difference is about $3.40, which isn't worth an engineering afternoon. At 100,000 runs it is about $34 a month, still small, so quality should decide the choice before price does. This is list-price arithmetic. Cache hits would lower DeepSeek's input cost, and reasoning tokens would raise its output cost, which is why your own measured usage should replace my assumptions. Of the two, classification is the safest step to move to the cheaper model. It produces 20 output tokens and has a checkable answer, so you can compare its labels against DeepSeek's on a sample before switching. Routing each step to its own model lets you keep DeepSeek only where the summaries need it, and that only pays if splitting the flow is free. In Activepieces it is: each step can call its own model through the DeepSeek integration, and the flow is metered at 1 credit per run, so the ten-step version costs what the two-step workaround did. Zapier bills per task and Make per module, so on those a finer split raises the automation bill and charges the builder for the step-by-step design that makes the routing worthwhile. ## Watch for hidden costs and limits Past the rate card, six things push a DeepSeek bill higher: reasoning tokens, long outputs, retries, prepaid balance rules, dynamic rate limits and data-location constraints. - Reasoning tokens (documented): DeepSeek bills the model's chain of thought as output, so cap output length and use the non-reasoning model for steps that don't need it. - Long outputs (documented): DeepSeek bills output at the higher rate, so instruct the model to answer briefly and set a maximum. - Retries (reported): a failed call re-sent in full pays for the prompt again, so use backoff and a retry limit. - Prepaid balance (documented): money is held in advance, so top up in small, planned amounts. - Dynamic rate limits and peak errors (limits documented, errors reported): throughput varies with load, so queue work and avoid synchronous dependence. - Data location (documented in the privacy policy): check it before sending customer data. ### Reasoning tokens and output length Log output tokens per step for a week, then price from the measured average. ### Rate limits and peak-hour errors Users report slow responses and errors at busy hours. Build in backoff, cap attempts, and run bulk jobs off-peak. ### DeepSeek API prepaid balance and credit terms Because DeepSeek is prepaid, you fund the account before spending. Read the current terms on whether granted credit expires and how it treats topped-up funds, and ask whether checkout adds tax. I wouldn't assume either. ### DeepSeek API data location and privacy in China Before you send customer data, note that DeepSeek's privacy policy states that it processes and stores **data on servers in China**. If your emails or tickets contain personal data, that's a legal question before it's a cost one. ## DeepSeek versus OpenAI, Anthropic and Google on price On a typical short task, DeepSeek V4-Pro sits within a factor of four of Google's Gemini 4 Argon, so the lead isn't uniform, even though it is the cheapest flagship of the four by a wide margin. ### Per-million-token price comparison In the table, input and output prices per million tokens come from [OpenAI's pricing page](https://openai.com/api/pricing/). It also gives what 1,000 runs cost if each run sends 1,000 tokens in and gets 200 back. That's 1M input tokens and 0.2M output tokens in total, roughly an email summary. | Model | Input $/1M | Output $/1M | 1,000-run cost | |---|---|---|---| | DeepSeek V4-Pro | 0.66 | 1.98 | $1.06 | | Gemini 4 Argon | 2 | 10 | $4.00 | | GPT-6 Astra | 10 | 50 | $20.00 | | Claude Fable 5.1 | 10 | 50 | $20.00 | DeepSeek V4-Pro charges 0.66 in and 1.98 out, so **output costs three times input**. Gemini 4 Argon charges 2 and 10. GPT-6 Astra and Claude Fable 5.1 both charge 10 and 50. For those three, output costs five times input, so a chatty model or a long reasoning trace hits their bills hardest. ![Input and output prices for flagship models](https://ap-marketing-media.fra1.cdn.digitaloceanspaces.com/uploads/31cae140-a513-48dc-9ea8-f1b1339290d3/deepseek-api-pricing-2026-cost-per-1m-tokens-dum-8cff609e.svg "Source: OpenAI") On V4-Pro the same 1,000 runs cost about $1, and on Astra or Fable $20, roughly nineteen times more. ### Where DeepSeek's lead holds and where it shrinks Against Astra and Fable the lead is largest. Input is about fifteen times pricier there and output about twenty-five times, and no cache discount closes a gap that size. Against Gemini 4 Argon, the gap narrows to under four times on this workload. On V4-Pro the same 1,000 runs cost about $1, and on Astra or Fable $20, roughly nineteen times more, making vendor selection the dominant cost factor. Then there are the budget tiers: DeepSeek's Flash model, GPT-6 Luna and Claude Haiku 5.5 are priced for exactly this kind of high-volume step. [Anthropic's model overview](https://docs.claude.com/en/docs/about-claude/models/overview) positions Haiku 5.5 as fast and cheap for high-volume work. Against those, a flagship price comparison is the wrong test. Price the task you actually run on the budget tier first. DeepSeek is also the wrong pick when your task fails on its quality, or when you need a capability its Pro model lacks. V4-Pro has no vision, so image-heavy workloads need a different model however low the rate. ## What Activepieces does about this The fix this article keeps arriving at is to measure your own workload, and that is easier when the flow is visible. Activepieces keeps a run history with each step's input and output, so the usage block DeepSeek returns on a real call, with its cache-hit and reasoning counts, is on screen next to the prompt that produced it. A few days of real runs give you the averages to replace my 1,500-in, 200-out assumptions. Cache hits depend on prompt order, and in a flow you can see that order. Keep the fixed instructions in the first part of the prompt, map the email or ticket text in last, and check the next runs for the hit count to move. Activepieces is open source under the MIT licence, so you can self-host it and run these checks on your own infrastructure, which matters given DeepSeek's data-location terms. The cost of the model calls is still the DeepSeek bill, and your measured figures are what set it. ## Frequently asked questions Vision isn't the only gap the rate card leaves, and five smaller questions decide whether the first invoice surprises you. ### Is the DeepSeek API free to use? No, the API is metered from the first token you send. The free product is the web chat, which is a separate thing with no bearing on API billing. ### Does DeepSeek offer a monthly subscription? Not as far as the API is concerned. You pay for usage, so there's no flat fee to amortise and no seat count to argue about. Nothing is bundled, so a runaway loop costs real money. ### Does DeepSeek charge for failed or cached requests? Billed, yes, but at the cheaper cache-hit input rate, which is the whole reason cache behaviour belongs in your estimate. Failed requests are less clear-cut. A call rejected before any processing should cost nothing, but one that generates output and then times out on your side may still be counted. ### Can I set a spending cap on the DeepSeek API? When your prepaid balance runs out, requests stop, and that is the practical cap. It's a blunt control, not a budget feature with alerts, so put your own guardrails in front of it. Count tokens per run, set a maximum output length on every call, and log reasoning tokens separately, because those are the ones that inflate a bill without anyone noticing. ### Are DeepSeek prices different at off-peak hours? DeepSeek now publishes time-of-day rates, but prices may change, so I wouldn't build a forecast on them. Whether a discount applies today, and whether it is applied automatically or needs action from you, should be read off the current pricing page. If you can defer batch work to a quiet window, price it both ways and keep the cheaper figure only if the page confirms it. ## Related reading - [Muse Wonder API Pricing: Rates per Million Tokens](https://www.activepieces.com/blog/muse-wonder-api-pricing-for-every-tier-and-workload) - [Gemini 4 Argon API Pricing: Tokens and Search (2026)](https://www.activepieces.com/blog/gemini-4-argon-api-pricing-tokens-and-search-2026) - [Bolt.new Pricing 2026: Plans, Tokens, and Costs Explained](https://www.activepieces.com/blog/boltnew-pricing-2026-plans-tokens-and-costs-explained) ## References - [GetAPIPulse](https://www.getapipulse.com/top-10-cheapest-llm-apis.html) - [DeepSeek API Docs](https://api-docs.deepseek.com/quick_start/pricing) - [OpenAI](https://openai.com/api/pricing/)