AI Tools

What does DeepSeek V4.1 Flash cost, and what happens to V4 Pro?

Published 2026-09-11 · Researched and price-checked on official sources

DeepSeek V4.1 Flash pricing starts at $0.15 per million cache-miss input tokens and $0.60 per million output tokens off-peak, under a quarter of V4 Pro's cache-miss rate. From noon Beijing time on September 14, 2026, every deepseek-v4-pro request is routed to V4.1 Flash and billed at the Flash price.

Key facts

What matters

deepseek-flash, the API name for DeepSeek-V4.1-Flash, bills $0.15 per 1M cache-miss input tokens and $0.60 per 1M output tokens off-peak; peak rates are exactly double.
V4 Pro's published card is $0.66 cache-miss input and $1.98 output off-peak, so V4.1 Flash costs 4.4x less on uncached input and 3.3x less on output.
From 12:00 Beijing time on September 14, 2026, requests to deepseek-v4-pro are routed to V4.1 Flash and billed at the Flash price until V4.1 Pro ships.
DeepSeek's own model card has V4.1 Flash beating V4 Pro on Terminal-Bench 2.1 (90.6 vs 87.9), DeepSWE v1.1 (74.2 vs 62.7) and AutomationBench (54.8 vs 31.8), but losing on HLE (36.8 vs 42.7) and SimpleQA world knowledge (42.3 vs 55.2).
The weights are open: deepseek-ai/DeepSeek-V4.1-Flash on Hugging Face is MIT-licensed, with a 552B-parameter backbone that activates 8B during prefill and 16B during decode.

What DeepSeek V4.1 Flash actually is

DeepSeek-V4.1-Flash is a multimodal Mixture-of-Experts model that DeepSeek put on its public API on September 10, 2026, according to the official change log. It is the smallest model in the new V4.1 architecture family and processes images as well as text, which separates it from the text-only models it replaces.

You call it as deepseek-flash. The legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp still return a response, but the models behind them were retired: DeepSeek serves those requests with V4.1 Flash and bills the Flash price. The OpenAI-format and Anthropic-format base URLs (api.deepseek.com and api.deepseek.com/anthropic) are unchanged, so no client rewrite is needed beyond the model string.

The model card on Hugging Face describes a 552B-parameter backbone that activates 8B parameters per token during prefill and 16B during decode, a 1M-token context window, and 45T training tokens. Its engineering headline is KV cache compression: the card puts the global cache at 890 bytes per token, roughly a quarter of DeepSeek-V4-Flash. The repository and the weights are licensed under MIT, and the repo ships the DeepSeek_V41_Tech_Report.pdf alongside a reference prompt encoder.

The verified V4.1 Flash and V4 Pro rate card

Every price below is per million tokens, read from DeepSeek's Models & Pricing page on September 11, 2026. Off-peak rates are exactly half of peak rates. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday; every other hour, including all weekend, is off-peak.

Cache-hit input: $0.003 off-peak and $0.006 peak on V4.1 Flash, against $0.022 and $0.044 on V4 Pro.
Cache-miss input: $0.15 off-peak and $0.30 peak, against $0.66 and $1.32 on V4 Pro.
Output: $0.60 off-peak and $1.20 peak, against $1.98 and $3.96 on V4 Pro.
Concurrency limit: 2,500 requests for Flash against 500 for V4 Pro.

Because the discount holds on every row, the saving does not depend on your input-to-output mix: 4.4x on uncached input, 3.3x on output, 7.3x on cached input. The change log states the V4.1 release came with an API price reduction, and this is the card in force now.

V4-Pro-0813 still sits on the same pricing table next to deepseek-flash. The name remains purchasable today; the retirement is a scheduled routing change rather than a removed product.

Why DeepSeek thinks it can retire V4 Pro

DeepSeek's model card publishes both models in one table, and on agentic work the cheap tier now leads. Terminal-Bench 2.1: 90.6 for V4.1 Flash against 87.9 for V4 Pro. DeepSWE v1.1: 74.2 against 62.7. AutomationBench: 54.8 against 31.8. Agents' Last Exam: 31.8 against 25.7. CyberGym: 88.1 against 83.3. SEC-Bench Pro: 62.8 against 56.4. HLE with tools: 63.9 against 60.0. Codeforces rating: 3471 against 3348.

Those aren't just an internal tier shuffle. The same table puts V4.1 Flash above GPT-5.6 Sol on Terminal-Bench 2.1 (90.6 vs 88.8) and DeepSWE v1.1 (74.2 vs 73.0), and about level with Opus-5.0 on both (89.1 and 74.0), while Opus-5.0 stays clearly ahead on the harder Terminal-Bench 3.0 (43.3 vs 30.0).

Then read it next to the invoice. OpenAI lists GPT-5.6 Sol at $20 per 1M output tokens on its standard short-context tier, with promotional pricing running through November 21, 2026. DeepSeek sells output roughly 33 times cheaper off-peak from a model that edges Sol on two agentic benchmarks, which is the real reason V4 Pro is being shuffled out rather than kept as a premium tier.

Where V4 Pro still wins, and how those numbers were produced

The retirement is not a clean sweep, and DeepSeek's own table shows where Pro holds ground. World knowledge is the clearest, in the model card's base-model evaluation table: SimpleQA-Verified 55.2 for V4 Pro against 42.3 for V4.1 Flash, MultiLoKo 50.9 against 45.5, HLE 42.7 against 36.8 on the text-only subset, and LongBench-V2 51.5 against 45.2. If your product is a knowledge assistant over long reports rather than a coding agent, the model being retired is still the stronger one on the axis you care about.

One caveat covers every figure above: DeepSeek ran them. The model card states its instruct results use temperature 1.0 and top_p 0.95 at maximum reasoning effort, with code-agent benchmarks run in DeepSeek Harness Minimal mode on a 1M-token context, DeepSWE v1.1 in the mini-SWE harness and SEC-Bench Pro in the Claude Code harness. No third party has reproduced the set. The harness spread alone is worth remembering: DeepSWE v1.1 scores 74.2 in mini-SWE but 65.5 in OpenCode, so roughly nine points of that column is your scaffold, not your model.

What changes in your code on September 14, 2026

The line to put in your calendar is a pricing footnote. From 12:00 Beijing time on September 14, 2026, and until V4.1 Pro is released, all requests to deepseek-v4-pro are routed to V4.1 Flash and billed at the V4.1 Flash price. DeepSeek has published no date or price for V4.1 Pro.

Pinning deepseek-v4-pro no longer pins the model. The string keeps working while the weights behind it change, which is the kind of swap that quietly breaks prompt-tuned output.
Move the parameter to deepseek-flash first, so request logs and cost attribution reflect what you are actually running.
Image inputs now work on the cheap tier: the pricing table marks Vision as supported for V4.1 Flash and not supported for V4 Pro.
Effort is dialable per request. The model card documents a continuously controllable reasoning effort from 1 to 100, and the quick-start example still passes reasoning_effort high.
Self-hosting is real but not casual: MIT weights, an FP8 checkpoint and a 552B-parameter backbone put this in multi-GPU territory, so most teams stay on the API.

Who should switch now, and who should re-test first

Move now if you run high-volume extraction or classification, fan-out agent jobs, or image pipelines that had no cheap option in this family. The concurrency ceiling alone is an argument: 2,500 parallel requests against V4 Pro's 500. Anyone already pinned to deepseek-v4-pro moves on September 14 whether they plan to or not, and moving earlier means the switch shows up in your own metrics instead of in a production incident.

Re-test before you move if your prompts lean on world knowledge, long-context recall, or a frozen eval baseline. Run the same prompt set against deepseek-flash before the routing date; that comparison is the cheapest insurance available. Do it off-peak and it costs half the list price, which means weekday hours outside 01:00 to 04:00 UTC and 06:00 to 10:00 UTC, or any hour at the weekend.

If your only question was the invoice, the answer is deepseek-flash off-peak: $0.15 per million uncached input tokens and $0.60 per million output tokens, from the same base URL you already use.

At a glance

Feature	deepseek-flash (V4.1 Flash)	deepseek-v4-pro (V4-Pro-0813)
Cache-hit input, off-peak / peak (per 1M)	$0.003 / $0.006	$0.022 / $0.044
Cache-miss input, off-peak / peak	$0.15 / $0.30	$0.66 / $1.32
Output, off-peak / peak	$0.60 / $1.20	$1.98 / $3.96
Concurrency limit	2,500	500
Vision input	Supported	Not supported
Context / max output	1M / 384K	1M / 384K
Terminal-Bench 2.1	90.6	87.9
DeepSWE v1.1	74.2	62.7
SimpleQA-Verified (world knowledge)	42.3	55.2

DeepSeek V4.1 Flash vs V4 Pro: published API facts, verified September 11, 2026

FAQ

Is DeepSeek V4.1 Flash cheaper than V4 Pro?

Yes, on every row of the published table. Off-peak, V4.1 Flash is $0.15 per 1M cache-miss input tokens and $0.60 per 1M output tokens against $0.66 and $1.98 for V4 Pro, and peak rates are exactly double those figures. That makes it 4.4x cheaper on uncached input, 3.3x on output and 7.3x on cached input, verified on DeepSeek's Models & Pricing page on September 11, 2026.

What happens to deepseek-v4-pro on September 14, 2026?

From 12:00 Beijing time that day, requests to the name deepseek-v4-pro are routed to V4.1 Flash and billed at the Flash price. DeepSeek has not published a release date or price for V4.1 Pro, so the routing stays in place until it launches.

Can I keep calling deepseek-v4-flash?

The name still answers, but the underlying model is retired: DeepSeek serves those requests with V4.1 Flash and bills the Flash price, and the same applies to deepseek-v4-flash-vision-exp. Switching the parameter to deepseek-flash keeps your logs and cost reports accurate.