DeepSeek Input vs Output Token Pricing, Explained

DeepSeek charges twice as much for output tokens as for input on V4-Flash, and roughly four times as much on the R1 reasoning model. That ratio, not the headline input price, is what actually sets your bill for any chat or agent workload. We confirmed it on a live call where a tiny 41-token prompt still spent 79% of its cost on the 75 tokens the model wrote back.

Most pricing pages quote a single "from $0.14" number and move on. The problem is that almost no real workload is input-only. The moment your model writes anything back, the output rate takes over, and on a chatty workload it is the number that decides whether your monthly bill is $200 or $600.

DeepSeek input vs output token pricing (verified 2026-05)

DeepSeek prices the two directions separately, and the multiplier between them shifts by model. That is why a price quote with no model attached tells you almost nothing. According to DeepSeek API Docs, each tier publishes a distinct input and output rate, so the full rate card is the only honest way to quote it.

Model	Input ($/1M)	Output ($/1M)	Output-to-input multiplier
DeepSeek V4-Flash	$0.14	$0.28	2×
DeepSeek V4-Pro (promo, to 2026-05-31)	$0.435	$0.87	2×
DeepSeek R1	$0.55	$2.19	~4×

Sources: DeepSeek API Docs – Pricing and DeepSeek API Docs – Pricing Details USD, verified 2026-05-13.

One naming trap worth flagging before you budget: when we called the deepseek-chat alias directly, the API returned deepseek-v4-flash as the actual model. The "legacy V3 at $0.27 input / $1.10 output" figure that several pricing roundups still attach to deepseek-chat is stale. Call the alias today and you get V4-Flash at $0.14 / $0.28. Budget against the model you actually receive, not the one the old docs describe.

We watched output dominate a real bill (first-hand evidence)

The fastest way to see why the multiplier matters is to read a real invoice line, so we sent one. Our full run is in the evidence pack; here is the call that makes the point.

We sent a deliberately input-light prompt: a short system instruction plus a one-line question, 41 input tokens in total. The model replied with 75 tokens. On paper this looks balanced enough that input should hold its own. It did not.

Component	Tokens	Rate	Cost
Input	41	$0.14/M	$0.0000057
Output	75	$0.28/M	$0.0000210
Total	116	combined	$0.0000267

Output was 79% of the bill on a prompt we engineered to be input-light. The reason is simple once you see it. Even at a near 1:1 token split, the 2× output rate does the heavy lifting. On a realistic chat turn, where the model writes far more than the user types, output routinely runs 80% or more of the per-message cost. That is the number to design around.

Why the multiplier changes between models

The output-to-input multiplier is not a billing quirk; it tracks how each model generates text. DeepSeek V4-Flash is a short-form completion model that keeps output close to input in price, which gives chat workloads a predictable, gently sloping cost curve. DeepSeek R1 is a chain-of-thought reasoning model that produces long internal traces before its final answer, so each output token carries more compute and the price reflects that with a roughly 4× multiplier.

Output-token multiplier is the ratio of output token price to input token price for a given model. For V4-Flash it is 2; for R1 it is about 4. The figure compounds with your traffic shape. A workload running one input token for every output token spends two-thirds of its budget on output with V4-Flash, but four-fifths of it with R1. Switch models without re-checking the multiplier and your cost model quietly breaks.

What this means when you pick a model

According to NxCode, the output-heavy traffic mix we measured is the norm for B2B chat rather than an edge case, which is exactly why the multiplier deserves more attention than the headline input rate. The practical move is to treat input price as the floor and the output multiplier as the real driver.

For most conversational and summarization work, V4-Flash and its 2× multiplier is the right default. Reach for R1 only when chain-of-thought reasoning measurably lifts accuracy on your own eval, and when you do, budget for the 4× output rate on those calls specifically. According to DeployBase, the most under-used cost lever in DeepSeek deployments is exactly this: route the reasoning-hungry queries to R1 and keep everything else on V4-Flash, rather than running one model for the whole workload. A blended routing strategy beats a single-model default almost every time output volume is high.

Choose this if you are budgeting a new workload

Assume the 2× output multiplier and price against output volume, not input, when the task is conversational or short-form. Sample a real day of traffic, count output tokens separately from input, and multiply each by its own rate before you trust any "cost per call" estimate. If your eval justifies R1 on a slice of queries, carve that slice out and budget it at the 4× rate on its own line. And ignore any quote that gives you one blended "per token" figure without naming the model and splitting the two directions, because that number cannot survive contact with a real workload.

FAQ

Why is DeepSeek's output token price higher than its input price? Generating output is more compute-intensive than reading input, because the model produces output one token at a time while input arrives all at once. DeepSeek prices V4-Flash output at $0.28 against $0.14 input as of May 2026, a 2× gap.

What is the output-to-input ratio for DeepSeek V4-Flash? Exactly 2×. Input is $0.14 per million tokens and output is $0.28. The same 2× ratio holds for V4-Pro at both its promo and list rates.

Does output really dominate the bill even on short prompts? Yes. In our live test a 41-token input prompt drew a 75-token reply, and output still accounted for 79% of the $0.0000267 total. On realistic chat turns where replies are longer than prompts, output's share climbs past 80%.

Is DeepSeek R1 always more expensive than V4-Flash? Per token, yes. R1 input is roughly 4× V4-Flash input ($0.55 vs $0.14) and R1 output is about 8× ($2.19 vs $0.28). R1 earns its premium only when reasoning quality on your eval improves enough to justify the spend.

This is part of the DeepSeek API pricing hub. For the cache mechanics that change the input side of this math, see the cache hit discount article.

Author: Kevin Fan, Customer Success Manager at China LLM Directory, specializing in Chinese LLM ecosystem pricing. Last verified: 2026-05-13.