DeepSeek V4-Flash vs V4-Pro

DeepSeek's API exposes only two real model names, V4-Flash and V4-Pro, and for most workloads V4-Flash wins: V4-Pro costs about 3x more and is materially slower, so reach for it only when your own eval shows a quality lift worth paying for. The two names are the whole menu. Everything else you may have read about, deepseek-chat, deepseek-reasoner, even deepseek-r1, either resolves to V4-Flash or no longer exists as a separate product. This article walks through the steady-state price gap, the latency gap we measured ourselves, and a simple decision rule.

DeepSeek V4-Flash vs V4-Pro pricing (verified 2026-06)

The two models sit at different price points, and the gap is the first thing to settle before you read anything into the "Pro" label. According to DeepSeek API Docs, V4-Flash bills at $0.14 per million input tokens and $0.28 per million output. V4-Pro bills at $0.435 input and $0.87 output, with cache-hit input at $0.003625. Those are the standing rates, re-verified against the official docs on 2026-06-25, with no countdown clock attached to either model.

Model	Input ($/1M)	Output ($/1M)	Cache-hit input ($/1M)
V4-Flash	$0.14	$0.28	$0.0028
V4-Pro	$0.435	$0.87	$0.003625

So the real multiplier is roughly 3x on both input and output: $0.435 against $0.14, $0.87 against $0.28. According to DeepSeek API Docs – Pricing Details, these are the live per-model rates, and that 3x is the number to build a budget on. A note on history, because the question comes up: V4-Pro was originally announced with a nominal $1.74 / $3.48 list price that a launch discount sat on top of, and many older write-ups still warn that the rate "reverts" to that figure. It never did. The $1.74 / $3.48 number never took effect, so treat $0.435 / $0.87 as what you actually pay and the 3x gap as the standing fact, full stop.

What the names actually resolve to

This is the trap. The official API accepts exactly two model strings, and the friendly aliases you see in older tutorials quietly map onto them. According to DeepSeek API Docs, the supported names are deepseek-v4-pro and deepseek-v4-flash; request anything else and you may get an error or a silent redirect.

We confirmed the redirect directly. When we called the endpoint with deepseek-chat, the response's served-model field came back as deepseek-v4-flash. We re-ran it with deepseek-reasoner and got the same answer: deepseek-v4-flash again, this time with reasoning tokens in the usage object. So both aliases bill at the V4-Flash rate of $0.14 / $0.28, regardless of which name your SDK was configured with years ago. The practical upshot: if your code still says deepseek-chat, you are already on V4-Flash and paying the V4-Flash price, whether you meant to or not.

deepseek-r1 is a separate story. Requesting it returns an HTTP 400 telling you the only supported names are V4-Pro and V4-Flash. R1 is no longer a purchasable model on the official API, so the $0.55 / $2.19 figure some price lists still carry for it is stale. Reasoning behavior now lives inside V4-Flash via the deepseek-reasoner alias, not in a separate SKU.

First-hand evidence: the latency gap we measured (verified 2026-06)

Documentation gives you prices, not response times, so we timed both models on the live endpoint. The headline is that V4-Flash earns its name. In our small-call latency runs, V4-Flash returned in about 0.7 seconds; we logged a 0.733s response on a 557-token cache test, the cleanest single read in that batch. On the same official endpoint, our V4-Pro call took 2.64 seconds. The cluster-wide aggregate we recorded across V4-Pro runs lands at about 2.13 seconds, which is the figure to plan around.

That is the gap that matters for anything user-facing. A chat box, an autocomplete, an agent loop that makes several sequential calls, all of them feel the difference between roughly 0.7s and roughly 2.1s on every turn. V4-Pro is not slow because it is broken; it is slower because it does more thinking per token. You are trading wall-clock time for whatever reasoning lift the heavier model gives you, and that trade only pays off if the lift is real for your task.

We also ran the same coding prompt through both DeepSeek and two Western comparators to sanity-check the speed story. Asking for an iterative Fibonacci function ("code only"), V4-Flash produced 120 output tokens from a 27-token prompt in 2.663 seconds. For reference, GPT-4o answered the same prompt in 2.437 seconds at a billed cost of $0.000795, and Claude Sonnet 4.6 took 3.274 seconds at $0.001344. Both comparators were called via OpenRouter, not their first-party APIs, so treat those as routed numbers; the DeepSeek call was on the official endpoint. The point is not that DeepSeek is fastest, it is that V4-Flash sits right in the same response-time band as far pricier models while charging a fraction of the token rate.

The decision: when to actually pay for V4-Pro

Here is the rule we would give a teammate. Default to V4-Flash. It is the model your deepseek-chat alias already serves, it answers in roughly a third of the time, and at $0.14 / $0.28 it is cheap enough that cost rarely forces your hand. Move to V4-Pro only when you have run your own evaluation on your own prompts and seen a quality lift that justifies both the latency hit and the 3x price.

Choose V4-Flash for chat, autocomplete, classification, high-volume batch work, and any latency-sensitive path. It is the cost and speed floor.
Choose V4-Pro when an eval shows it clearing a quality bar V4-Flash cannot, on hard reasoning, multi-step planning, or tasks where a wrong answer is expensive.
Run the eval before you switch. V4-Pro is about 3x the token rate and roughly 3x the latency. Size your bill and your response budget on those standing numbers, not on the "Pro" label.

If your eval comes back a wash, that is your answer: stay on V4-Flash. Paying 3x for output you cannot distinguish in testing is not an upgrade, it is a leak. The whole reason to test before switching is that "Pro" in the name is marketing, while the latency and the price are facts you can measure. For the full rate card across cache tiers, the DeepSeek API pricing hub lays out every number in one place.

FAQ

Are V4-Flash and V4-Pro really the only DeepSeek models? On the official api.deepseek.com endpoint, yes. The API accepts only deepseek-v4-flash and deepseek-v4-pro. The familiar deepseek-chat and deepseek-reasoner are aliases that both resolve to V4-Flash, and deepseek-r1 now returns an HTTP 400 because it is no longer a separate purchasable model.

How much faster is V4-Flash than V4-Pro? In our tests, V4-Flash returned small calls in about 0.7 seconds (a logged 0.733s on one read), while V4-Pro took 2.64 seconds on a comparable call, with an aggregate near 2.13s. That is roughly a 3x speed advantage for V4-Flash on every turn, which is why it suits chat and other latency-sensitive paths.

How much more does V4-Pro cost than V4-Flash? About 3x on both sides of the bill. V4-Flash is $0.14 input / $0.28 output; V4-Pro is $0.435 input / $0.87 output, re-verified against the official docs on 2026-06-25. An older $1.74 / $3.48 nominal list figure circulated at launch but never took effect, so it is not what you pay today.

Which model does deepseek-chat use? It uses V4-Flash. When we called the endpoint with deepseek-chat, the served model came back as deepseek-v4-flash, billing at $0.14 / $0.28. If your code still references the old alias, you are already on V4-Flash and paying its rate.

Is V4-Pro's reasoning worth the premium? Only your eval can answer that. V4-Pro spends more compute per token, which is why it is slower, but the quality lift is task-dependent. Run both on your own prompts; if the outputs are indistinguishable, the 3x cost and slower response make V4-Flash the clear pick.

This is part of the DeepSeek API pricing hub. For help reading the bill, see the input vs output pricing breakdown, and to cut the input side, the cache discount explainer.

Author: Kevin Fan, Customer Success Manager at China LLM Directory, specializing in Chinese LLM ecosystem pricing. Last verified: 2026-06-25.

DeepSeek's API exposes only two real model names, V4-Flash and V4-Pro, and for most workloads V4-Flash wins: V4-Pro costs about 3x more and is materially…

Written by Fan Chuanyu