VERIFIED PRICETHINKING FLAGSHIP1MTEXT MODELPROMPT CACHING
DeepSeek V4 Pro API Pricing
DeepSeek bills by time of day from 16:00 UTC on August 16, 2026. Peak hours are 01:00-04:00 and 06:00-10:00 UTC; every other hour is off-peak at exactly half the peak rate. The tiles below carry the peak rate, because DeepSeek describes off-peak as "half the peak rates" - peak is the list price and off-peak is the discount. DeepSeek V4 Pro is the DeepSeek thinking flagship, and at peak it is $1.32/M input, $3.96/M output and $0.044/M cached input; off-peak it is $0.66 / $1.98 / $0.022. That replaced a flat $0.435 / $0.87 card which had itself survived an announced reversion to $1.74/$3.48 that DeepSeek never applied - so this is the first rise that actually landed.
Input - per 1M tokens
$1.32/M
Off-peak $0.66/M peak
Output - per 1M tokens
$3.96/M
Off-peak $1.98/M peak
Cached input
$0.044/M
Off-peak $0.022/M peak
Effective - agentic blend
$0.569/M
92/8 split - 82% cache
§ 01 / TERMINAL
DeepSeek V4 Pro API cost calculator.
Calculator pre-loaded with DeepSeek V4 Pro PEAK rates — halve every result for an off-peak run, which covers 17 of every 24 hours. Tweak spend, output mix, or cache hit rate to compare this model with nearby alternatives.
$ /mo
Workload split
Prompt cache hit rate
Tokens you can process
—
Words equivalent (English)
—
Effective rate
—
§ 02 / SCENARIOS
DeepSeek V4 Pro cost on real workloads.
AGENT LOOP
Repo-wide bug fix
$0.072/task
LONG DOC
Reading 100-page contracts
$0.081/doc
RAG SUPPORT
Support agent ticket triage
$0.00615/ticket
ASSISTANT TURN
Research planning turn
$0.013/turn
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 344
Words 56
Tokens (estimated) 65 tokens
Cost as input · uncached $0.00009 USD
Cost as output · uncached $0.00026 USD
Cost as cached input $0.000003 USD
| Model | Input /M | Output /M | Effective blended | Context | Best for |
|---|---|---|---|---|---|
| DeepSeek V4 Pro Current | $1.32 cache $0.044 | $3.96 | $0.569 agentic 92/8 | 1M | Low-cost reasoning |
| DeepSeek V4.1 Flash | $0.30 cache $0.006 | $1.20 | $0.15 −3.8× | 1M | Cheap DeepSeek workloads, now with vision |
| GPT-6.1 Terra | $1.00 cache $0.05 | $5.00 | $0.645 +13% | 1.05M | OpenAI's fast, efficient GPT-6 tier |
| Claude Sonnet 5.5 | $2.00 cache $0.10 | $10.00 | $1.29 +2.3× | 1M | Anthropic's current Sonnet |
| Gemini 3.8 Flash | $0.75 cache $0.075 | $3.75 | $0.481 −15% · promo | 1M | Mainstream multimodal work |
| Grok 4.3 | $1.25 cache $0.20 | $2.50 | $0.558 same | 1M | Grok long-context agents |
By comparison
DeepSeek V4 Pro has published scores on 16 benchmarks we track; its best placing is #2 of 8 on NL2Repo-Bench. 2 of these scores come from an independent leaderboard; the rest are the lab's own reports.
| Benchmark | Score | Rank | Leader | Source |
|---|---|---|---|---|
| NL2Repo-Bench Coding | 61.5% score | #2 of 8 | DeepSeek V4.1 Flash · 68.2% | Vendor-reported DeepSeek |
| ARC-AGI-1 Reasoning | 90.5% ARC Prize verified, best result | #31 of 99 | 6 tied · 98.5% | Official leaderboard ARC Prize Foundation |
| ARC-AGI-2 Reasoning | 61.3% ARC Prize verified, best result | #33 of 105 | GPT-6 Astra · 95% | Official leaderboard ARC Prize Foundation |
| Toolathlon Agents and computer use | 74.1% as reported | #3 of 9 | MiMo-V2.6-Pro · 76.9% | Vendor-reported DeepSeek |
| Terminal-Bench 2 Agents and computer use · Terminal-Bench 2.1 | 87.9% pass@1 | #6 of 17 | DeepSeek V4.1 Flash · 90.6% | Vendor-reported DeepSeek |
| AutomationBench Agents and computer use · Main set (as reported) | 43.2% pass@1 | #9 of 22 | MiniMax M3.1 Flash · 62.8% | Vendor-reported DeepSeek |
| Humanity's Last Exam Expert knowledge · With tools | 60% pass@1, with tools | #9 of 20 | Claude Fable 5.2 · 74.2% | Vendor-reported DeepSeek |
| CyberGym Cybersecurity | 83.3% pass@1 | #5 of 8 | MiMo-V2.6-Flash · 95.1% | Vendor-reported DeepSeek |
| GPQA Diamond Expert knowledge | 92.4% pass@1 | #7 of 11 | GPT-6 Astra · 96% | Vendor-reported DeepSeek |
| Agents' Last Exam Agents and computer use · CLI split | 25.7% pass@1 | #4 of 6 | Kimi K3.1 Pro · 44.2% | Vendor-reported DeepSeek |
| DeepSWE Coding | 62.7% resolved rate | #18 of 27 | Gemini 4 Argon · 77.9% | Vendor-reported DeepSeek |
| Terminal-Bench 3.0 Agents and computer use | 11.8% pass@1 | #6 of 9 | DeepSeek V4.1 Flash · 30% | Vendor-reported DeepSeek |
| SEC-bench Pro Cybersecurity | 56.4% pass@1 | #6 of 8 | Gemini 4 Carbon · 88.2% | Vendor-reported DeepSeek |
| Terminal-Bench 4.0 Agents and computer use | 12.4% pass@1 | #29 of 34 | Kimi K3.1 Pro · 65.7% | Vendor-reported DeepSeek |
| ProgramBench Coding | 15.5% almost@1 | #6 of 7 | Kimi K3.1 Pro · 31.5% | Vendor-reported DeepSeek |
| ExploitGym Cybersecurity | 5.4% pass@1 | #8 of 9 | Gemini 4 Carbon · 54.8% | Vendor-reported DeepSeek |
DeepSeek V4 Pro API pricing FAQ.
Short answers for teams checking DeepSeek V4 Pro pricing, billing behaviour, and how it differs from V4 Flash.
Q · 01 How much does DeepSeek V4 Pro cost? +
DeepSeek lists two tiers since August 16, 2026. At peak (01:00-04:00 and 06:00-10:00 UTC) it is $1.32/M cache-miss input, $0.044/M cache-hit input and $3.96/M output; off-peak — the other 17 hours — is exactly half: $0.66 / $0.022 / $1.98. The context window is 1M tokens either way. Under AI//COST's 92/8 agentic blend with 82% cache hits, the effective planning figure is $0.57/M.
Q · 02 Is this a promotional price that will expire? +
It arrived as one, and the expiry did not happen. DeepSeek launched V4 Pro at a 75% discount and said the price would revert to
$1.74/$3.48 after May 31, 2026. That date passed with no reprice; we re-verified the pricing page on August 16, 2026, when peak/off-peak billing replaced that card with $1.32/$3.96 and $0.66/$1.98 — this is what the vendor charges. We now treat it as the standard list rate rather than a countdown, and re-check it on every pass over DeepSeek's card — the date of the last one is in the byline below.Q · 03 Is a price change coming? +
None is announced. Through early August the page carried an unspecified warning that DeepSeek planned to raise overall API pricing in the near future, with a significant increase expected. That increase landed on August 16, 2026 as the peak/off-peak card above, and the warning has since been removed — as of August 23, 2026 the page carries only the standing note that prices may vary. Nothing is scheduled, so there is no date to watch. Keep in mind that this vendor's previous announced change, the May 31 reversion on this very model, never happened: on DeepSeek an announcement is a signal, not a schedule.
Q · 04 How does it differ from V4 Flash? +
V4 Pro is the 1.6T-parameter MoE (49B active) against V4 Flash's 284B (13B active), and it runs about 3× the effective rate —
$0.57/M against $0.19/M at peak — for the same 1M context and the same 384K max output. Both support thinking and non-thinking modes at one price. Buy Pro for the harder reasoning, not for the context window.Q · 05 Does thinking mode cost extra? +
No. One rate card covers both non-thinking and thinking modes, and thinking is the default. The mode is free; the tokens it generates are billed at the normal output rate, so a reasoning-heavy request still costs more than a terse one.
Q · 06 Can I call it through the OpenAI Responses API? +
Yes, as of August 23, 2026. DeepSeek's feature table now marks the Responses API as supported on all three models, V4 Pro included — the gap that existed when this page was first written has closed. The OpenAI-compatible Chat Completions endpoint and the Anthropic-format endpoint work with V4 Pro too. The remaining difference from V4 Flash is throughput, not protocol: Pro is limited to 500 concurrent requests against Flash's 2500.
Q · 07 Do deepseek-chat and deepseek-reasoner still work? +
No. Both ids were fully retired after July 24, 2026, 15:59 UTC. Reasoning traffic that used
deepseek-reasoner now goes to deepseek-v4-pro or deepseek-v4-flash with thinking mode on.