Last verified
RETIRED SEP 10, 2026ID NOW SERVES V4.1 FLASH1M CONTEXTTHINKING SAME PRICECACHE HIT 31x CHEAPER

DeepSeek V4 Flash API Pricing

Retired on September 10, 2026. DeepSeek's change log says V4 Flash "has been retired" and that the name is temporarily routed to V4.1 Flash, billed at that model's card — $0.30 / $0.006 / $1.20 at peak, lower on all three lines than anything on this page. So calls still answer and nobody is overcharged, but temporarily is the vendor's own word: the routing can stop without a second notice, and code naming deepseek-v4-flash is pointing at a model that no longer exists. The figures below are frozen at what V4 Flash charged while it was sold. From 16:00 UTC on August 16, 2026 it billed by time of day — peak 01:00-04:00 and 06:00-10:00 UTC, every other hour at exactly half — $0.44/M input, $1.32/M output and $0.014/M cached input at peak, $0.22 / $0.66 / $0.007 off-peak. That card replaced a flat $0.14 / $0.28 / $0.0028 one and roughly quadrupled the effective blended rate at peak. Thinking and non-thinking modes billed alike, and from July 31, 2026 the id served checkpoint DeepSeek-V4-Flash-0731 at unchanged prices.

Input - per 1M tokens
$0.44/M
Off-peak $0.22/M peak
Output - per 1M tokens
$1.32/M
Off-peak $0.66/M peak
Cached input
$0.014/M
Off-peak $0.007/M peak
Effective - agentic blend
$0.189/M
92/8 split - 82% cache
§ 01 / TERMINAL

Run the numbers.

Calculator kept as a record, pre-loaded with the PEAK rates V4 Flash charged until it was retired — halve every result for an off-peak run. It no longer prices a live call: this id is served by DeepSeek V4.1 Flash today and billed at that model's lower card.

$ /mo
Workload split
Prompt cache hit rate
Tokens you can process
—
Words equivalent (English)
—
Effective rate
—
Open full calculator (all models · share URL · CSV) →
§ 02 / SCENARIOS

Real-world presets.

AGENT LOOP

Repo-wide bug fix

$0.024/task
120,000 in - 10,000 out~4,152 units/$100
LONG DOC

Reading 100-page contracts

$0.027/doc
180,000 in - 8,000 out~3,719 units/$100
RAG SUPPORT

Support agent ticket triage

$0.00205/ticket
8,000 in - 1,000 out~48,889 units/$100
ASSISTANT TURN

Research planning turn

$0.00427/turn
18,000 in - 2,000 out~23,406 units/$100
§ 03 / TOKENIZER

Paste text. See tokens. See cost.

What's a token →
Calibrated · measured on the vendor's tokenizer · 2026-06-10 Auto-counts as you type

Counts use a chars-per-token calibration measured on the vendor's own published tokenizer (deepseek-ai/DeepSeek-V3.2-Exp, 2026-06-10). English prose is typically within a few percent; code and non-Latin scripts tokenize heavier. For billing-exact counts use the vendor's count-tokens API.

Characters 444
Words 79
Tokens (estimated) 84 tokens
Cost as input · uncached $0.00004 USD
Cost as output · uncached $0.00011 USD
Cost as cached input $0.000001 USD
§ 04 / SHELF

Up against the shelf.

All models →
Model Input /M Output /M Effective blended Context Best for
DeepSeek V4 Flash Current $0.44 cache $0.014 $1.32 $0.189 agentic 92/8 1M What it charged while it was sold
DeepSeek V4.1 Flash $0.30 cache $0.006 $1.20 $0.15 −21% 1M The replacement, and it is cheaper
DeepSeek V4 Pro $1.32 cache $0.044 $3.96 $0.569 +3.0× 1M Harder reasoning, same 1M window
GPT-5.4 mini $0.75 cache $0.075 $4.50 $0.541 +2.9× 400K Teams already on OpenAI tooling
Claude Sonnet 4.6 $3.00 cache $0.30 $15.00 $2.05 +11× 1M Long agentic coding sessions
Gemini 3.8 Flash $0.75 cache $0.075 $3.75 $0.481 +2.5× · promo 1M Google stack, multimodal input
Grok 4.3 $1.25 cache $0.20 $2.50 $0.558 +3.0× 1M Grok long-context agents
§ 05 / DEEP LINKS

Specific scenarios.

All calculators →

By comparison

Compare DeepSeek → Compare Mistral → Compare Grok →

Frequently asked.

Short answers for teams checking DeepSeek V4 Flash pricing, billing behaviour, and migration choices.

Q · 01 Why are there two prices for DeepSeek V4 Flash? +
Because DeepSeek bills by the hour of day since 16:00 UTC on August 16, 2026. Peak hours are 01:00-04:00 and 06:00-10:00 UTC, Monday to Friday; every other hour is off-peak at half the peak rate. For this model the peak card is $0.44/M input on a cache miss, $0.014/M on a cache hit and $1.32/M output; off-peak is $0.22 / $0.007 / $0.66. The card it replaced was flat at $0.14 / $0.0028 / $0.28, so output rose 2.4x off-peak and 4.7x at peak. Re-verified on the vendor's page on August 23, 2026: the scheme is live and the rates are unchanged since the switch.
Q · 02 How much does DeepSeek V4 Flash cost? +
Nothing now — it was retired on September 10, 2026, and the id bills at V4.1 Flash rates instead. While it was sold, DeepSeek listed two tiers from August 16, 2026: at peak (01:00-04:00 and 06:00-10:00 UTC) $0.44/M cache-miss input, $0.014/M cache-hit input and $1.32/M output; off-peak, the other 17 hours, exactly half: $0.22 / $0.007 / $0.66. The context window was 1M tokens either way. Under AI//COST's 92/8 agentic blend with 82% cache hits that came to an effective $0.19/M — the model now answering to this id blends to $0.15/M.
Q · 03 Which model does deepseek-v4-flash actually serve? +
V4.1 Flash, since September 10, 2026. Until then it served DeepSeek-V4-Flash-0731, a checkpoint swapped in on July 31, 2026 without a price change. The V4.1 Flash release retired V4 Flash and, in DeepSeek's word, temporarily routed the name to the new model — so the id you call and the model you get have come apart. Billing follows the model rather than the name, so you are charged the lower card. Point new code at the replacement's own id, which is deepseek-flash and not deepseek-v4.1-flash; the V4.1 Flash page has the rates.
Q · 04 Is this page still worth reading? +
As a record, and for one live consequence. The rates here are frozen — nothing charges them any more. What is still true is the routing: code naming deepseek-v4-flash keeps working and is billed V4.1 Flash's lower card, so the failure mode is not a broken build or a surprise invoice, it is a silent model swap. Worth remembering how this vendor behaves. The price rise it warned about through early August arrived on August 16, 2026 as the peak/off-peak card above; the May 31 reversion it announced on V4 Pro never happened at all; and this retirement came with no deprecation window whatsoever. On DeepSeek an announcement is a signal rather than a schedule — and the absence of one guarantees nothing.
Q · 05 Do deepseek-chat and deepseek-reasoner still work? +
No. DeepSeek retired both ids after July 24, 2026, 15:59 UTC and they are now inaccessible. Non-thinking traffic that used deepseek-chat and reasoning traffic that used deepseek-reasoner both move to deepseek-v4-flash (or deepseek-v4-pro), switching between the two behaviours via thinking mode rather than via the model id.
Q · 06 Does thinking mode cost extra? +
No. The rate card publishes one set of prices per model, and it covers both non-thinking and thinking modes — thinking is the default. Note that thinking still produces tokens, so a reasoning-heavy request bills more than a terse one at the same rate; the mode is free, the tokens are not.
Q · 07 How cheap is the cache, really? +
A cache hit is $0.014/M against $0.44/M for a miss at peak — a 31x spread, and DeepSeek states cache storage itself is free. The ratio is identical off-peak, since both halve. That is why the effective blended figure lands near $0.19/M at peak rather than near the headline input rate. If your prompts do not repeat, budget the full miss rate instead — $0.44/M at peak, $0.22/M off-peak.
Q · 08 What is the output limit? +
DeepSeek lists a maximum output of 384K tokens for both V4 Flash and V4 Pro, inside the shared 1M-token context window.
Q · 09 When should I pay for V4 Pro instead? +
Read this as V4 Pro against the model that now answers here, V4.1 Flash: Pro runs about 3.8x the effective rate ($0.57/M against $0.15/M at peak) for the same 1M context, so the question is whether the harder reasoning earns that. Against V4 Flash as it was priced, the gap was 3x. The endpoint caveat that used to sit here is gone: as of August 23, 2026 DeepSeek's feature table marks the Responses API as supported on all chat models, V4 Pro included, alongside the Anthropic-format endpoint. Pro's concurrency limit is the remaining gap — 500 against 2500 on the Flash line.
Y
Reviewed by Yaroslav Vikhariev Founder · AI//COST · Prices checked by hand against api-docs.deepseek.com · Last verified Sep 19, 2026
Methodology Report a correction More by Y.V.