DeepSeek V4-Pro Pricing: $0.66/$1.98 per 1M Tokens

This page is a living rate card for the DeepSeek V4 family, not a news story. The numbers below were read directly off DeepSeek's official Models & Pricing page on 23 August 2026, and the price history further down explains how the rates got here — which is the part a bare pricing table can't tell you, and the part that predicts what happens next.

What does DeepSeek V4-Pro cost right now?

Since 16:00 UTC on 16 August 2026, all DeepSeek V4 models bill on a two-tier peak / off-peak schedule. Off-peak is exactly half of peak on every line item. These are the current published rates, per one million tokens, in USD:

Effective 16 August 2026, 16:00 UTC. Verified 23 August 2026 against api-docs.deepseek.com. All models: 1M-token context, 384K max output.

Three things about this table are easy to get wrong and worth stating plainly.

• There is no separate "V4-Pro Max" or subscription tier. Three model names exist on the API — deepseek-v4-pro, deepseek-v4-flash, and the experimental multimodal deepseek-v4-flash-vision-exp. That's the whole catalogue.
• Off-peak is not a discount you apply for. It's the default rate for most of the clock. You get it automatically by sending traffic outside the peak window.
• The concurrency limit, not a token-per-minute quota, is the throttle. V4-Pro allows 500 concurrent in-flight requests per account (and per user_id if you've had your quota raised); exceed it and you get an HTTP 429. DeepSeek publishes no RPM or TPM ceiling, and capacity expansion requests carry no extra cost.

When are DeepSeek's peak and off-peak hours?

Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday. Everything else — including all day Saturday and Sunday — is off-peak. That's 35 peak hours out of 168 in a week, so 79% of the week bills at the cheaper rate.

The weekday-only qualifier is the detail most third-party pricing pages drop, and it changes the arithmetic a lot. Here is what the peak window looks like in the timezones most teams actually work in:

Line the windows up against Beijing time and the design intent is obvious: peak is 09:00–12:00 and 14:00–18:00 in DeepSeek's home market, with the lunch hour carved out. This is capacity rationing for domestic daytime load, not a general price rise dressed up as one — DeepSeek told reporters the structure exists to "allocate resources more reasonably" and to push users to "schedule their tasks based on actual usage."

The practical consequence for a US or Latin American team: your interactive traffic is almost entirely off-peak already, so the effective V4-Pro price you'll see on an invoice is $0.66 / $1.98. For a European or Indian team, roughly half your working day lands in peak, and batch work should be shifted after 10:00 UTC. If your traffic is spread uniformly across the whole week, the blended rate works out to about $0.80 per million input and $2.39 per million output.

What is DeepSeek's cache-hit pricing, and how do you actually hit the cache?

Cache-hit input on V4-Pro is $0.022 per million tokens off-peak, $0.044 at peak — 30x cheaper than a cache miss. This is the single largest lever on a DeepSeek bill, and it's the line item that rose the most in August (more on that below).

Context caching is on by default for every account; there is no cache_control header to set and no opt-in. But V4's cache does not work like a naive prefix match, and the rules matter if you're trying to engineer a high hit rate:

• A request only hits the cache if it fully matches a persisted "cache prefix unit". Partial overlaps with a longer stored prefix do not count.
• Units are persisted at three moments: at the end of each user input and each model output (so multi-turn conversations cache naturally), when the system detects a common prefix shared across several requests, and at fixed token intervals inside very long inputs or outputs.
• The common-prefix detection means a shared document only becomes cacheable after two or more requests have used it. In DeepSeek's own example, a long financial report queried three different ways misses on the first two calls and hits on the third.
• Check your real hit rate with the prompt_cache_hit_tokens and prompt_cache_miss_tokens fields returned in the usage block of every response. If you are not logging these, you do not know what your bill is made of.

The design implication: front-load everything reusable — system prompts, RAG corpora, tool schemas, few-shot examples — into a stable prefix, and keep the variable part at the end. A workload that gets 80% of its input tokens from cache pays a blended input rate of about $0.15/M off-peak instead of $0.66/M.

When was DeepSeek V4-Pro released, and what is its knowledge cutoff?

V4-Pro was released on 24 April 2026 as part of the DeepSeek-V4 Preview launch, and reached general availability on 13 August 2026 as checkpoint DeepSeek-V4-Pro-0813. The model name on the API has never changed — deepseek-v4-pro always routes to the newest checkpoint.

On the knowledge cutoff: DeepSeek does not publish an official training-data cutoff for the V4 family. Neither the Hugging Face model card nor the API docs state one. Third-party model catalogues list April 2026, consistent with the 24 April release date, and that is the best available figure — but treat it as an inference, not a vendor statement. If your application depends on the model knowing about events after early 2026, test rather than assume, and note that DeepSeek did not say whether the August 0813 checkpoint extended the cutoff.

The full dated timeline, useful if you're auditing which version you were billed for:

Two operational notes fall out of that table. The legacy aliases are gone, not deprecated — if you have config still referencing deepseek-chat, it is already failing in production. And the GA checkpoint added reasoning_effort, which is a direct cost control: dropping from max to low on simple tasks cuts output tokens, and output is where the money goes.

Is DeepSeek V4-Pro cheaper than Claude Opus 5, GPT-5.6 and Gemini?

Yes, by a wide margin — smaller than it was in July, still large. Current published list rates for the frontier tier, per million tokens:

Verified 23 August 2026 from each vendor's official pricing page. GPT-5.6 Sol's $4/$20 is promotional pricing OpenAI says runs at least through 21 November 2026. Gemini 3.7 Flash sits at $0.75/$3.75 through 31 December 2026 and doubles on 1 January 2027.

The cache-read column is where the gap is widest and least discussed. Claude Opus 5 charges $0.50 per million cached input tokens; V4-Pro charges $0.022 off-peak. That's 22.7x. For any workload that re-reads a large fixed context — document Q&A, codebase-aware assistants, long agent transcripts — the cache line dominates the bill, and DeepSeek's is in a different order of magnitude.

The honest counterweight: V4-Pro is text-first (vision lives on the experimental Flash variant), and long-horizon agentic reliability across hundreds of chained tool calls is still where Opus 5 and GPT-5.6 Sol earn their premium. Price the workload, not the model. For a head-to-head on the open-weights side, our DeepSeek V4 vs Kimi K3 comparison goes through the benchmark deltas at these price points.

What does a real monthly bill actually look like?

Rate cards mislead because input and output volumes are never equal. Take a concrete, mid-sized workload: 50 million input tokens and 10 million output tokens per month — roughly a production coding assistant or document-processing pipeline serving a small team.

Read the top two rows together, because that comparison is entirely under your control. Getting the cache hit rate from 0% to 80% saves $25.52 a month on this workload — about as much as scheduling every request off-peak does. On a workload ten times this size, the same two levers are worth roughly $500 a month, and neither requires changing model or vendor.

If your volumes are large and the quality bar is "good enough at any frontier model" — bulk extraction, classification, summarisation, structured generation — V4-Flash at $0.22 / $0.66 does the same job for a third of V4-Pro's cost. Our V4-Flash deep dive covers where the quality actually diverges from Pro.

How did DeepSeek's pricing get here — and where is it going?

This is the part worth understanding, because DeepSeek has now run the same play twice in opposite directions, and the pattern is predictive.

V4-Pro launched on 24 April 2026 at $1.74 input / $3.48 output, with a 75% promotional discount attached and a stated expiry. In late May, rather than let the promo roll off, DeepSeek made the discounted number the list price: $0.435 input, $0.87 output, $0.003625 cached input, flat, all hours. There was no expiry date attached. For roughly three months that stood as the market's reference price — the first genuinely frontier-adjacent model to sit under a dollar per million output tokens.

On 13 August, DeepSeek announced that as of 16:00 UTC on 16 August the flat rate would be replaced by the peak/off-peak schedule. Measured against the flat May rates, the increases were:

• Input (cache miss): $0.435 → $0.66 off-peak (1.52x) or $1.32 peak (3.03x)
• Output: $0.87 → $1.98 off-peak (2.28x) or $3.96 peak (4.55x)
• Cached input: $0.003625 → $0.022 off-peak (6.07x) or $0.044 peak (12.1x, a 1,114% increase)

That cache-read number is why the coverage ran headlines about an 1,100% price rise. It is technically accurate and practically misleading: cached input is such a small fraction of most bills that even a 12x rise on it moves the total less than the 2.28x on output does. But for the specific workloads that had engineered themselves around near-free cache reads — long-context document pipelines especially — it was the single most disruptive line.

The important framing is that this was not a pure price rise. It shipped the same day as the V4-Pro GA checkpoint, with agent upgrades, configurable reasoning effort, and Codex-optimised Responses API support. DeepSeek attributed the structure to capacity constraints, saying it wanted to "allocate resources more reasonably." Read alongside the fact that peak hours map exactly onto Chinese office hours, the honest interpretation is: demand outran serving capacity, and rather than degrade latency or hard-cap concurrency, DeepSeek priced the scarce hours.

What that tells you about the next move:

1. The off-peak rate is the real rate, and it is defensible. DeepSeek did not raise the floor to squeeze margin — it raised the ceiling on hours it can't serve cheaply. A vendor optimising for margin would have raised both tiers uniformly.
2. Expect the pattern to repeat on the next model. Launch promo → promo becomes list → list gets restructured when capacity binds. Budget on the assumption that any DeepSeek launch price is a floor with a 12-week half-life, not a contract.
3. Cache economics are now a deliberate lever, not a giveaway. The 12x cache-read increase is the clearest signal that DeepSeek was subsidising long-context reuse and has stopped. Don't build a business model on any single vendor's cache being free.
4. Nobody matched the May floor and nobody has undercut the August one. GLM-5.3 at $1.40/$4.40 is the nearest competitor and is still above V4-Pro off-peak on both dimensions. The sub-dollar frontier output token was a moment, not a trend.

Does self-hosting beat the API at these prices?

Rarely, but the August increase moved the break-even meaningfully closer than it was in July. The weights are open — MIT licence, deepseek-ai/DeepSeek-V4-Pro on Hugging Face — so the option is genuinely available, not theoretical.

The blocker is scale. V4-Pro is 1.6T total parameters with 49B active per token; a full-precision deployment needs multiple 8-GPU nodes, and even aggressive quantisation trades away long-context quality, which is the whole point of a 1M-token model. At $0.66 / $1.98 off-peak you can buy an enormous number of tokens before you amortise a single node, and you get the 0813 checkpoint upgrades for free. Self-hosting makes sense when data residency is a hard requirement, when you already own idle GPUs, or when you need the model offline — not on price alone. The per-quantisation numbers are in our DeepSeek V4 VRAM and GPU requirements guide.

How do you call V4-Pro at the current rate?

There is nothing to configure to get the pricing — the rate follows the clock. The two integration details that still trip people up:

from openai import OpenAI

client = OpenAI(
    api_key=os.environ["DEEPSEEK_API_KEY"],
    base_url="https://api.deepseek.com",   # no /v1 suffix — the SDK appends it
)

resp = client.chat.completions.create(
    model="deepseek-v4-pro",
    messages=[{"role": "user", "content": "Hi"}],
    reasoning_effort="low",                # low | high | max — the cheapest cost lever
    extra_body={"thinking": {"type": "enabled"}},
)

Baking /v1 into the base URL is the most common cause of unexplained 404s on a valid key; older V3.2-era tutorials all did it. An Anthropic-compatible endpoint also exists at https://api.deepseek.com/anthropic, which makes Claude Code and other Messages API clients a drop-in swap — set ANTHROPIC_BASE_URL and ANTHROPIC_MODEL=deepseek-v4-pro. Caching is automatic there too; cache_control blocks are ignored rather than honoured.

What should you actually do about it?

Three decisions, in order of how much money they move:

1. Measure your peak-hour exposure before optimising anything else. If you're a US-hours team, you're already paying off-peak on essentially all interactive traffic and there is nothing to fix. If you're in Europe or India, move every batch, eval, and backfill job outside 01:00–04:00 and 06:00–10:00 UTC — that's a 50% cut on the shifted portion for zero code change.
2. Instrument prompt_cache_hit_tokens and act on it. At a 30:1 ratio between miss and hit, cache design is worth more than model selection for most long-context workloads. Stable prefix first, variable content last.
3. Right-size the model and the reasoning effort. V4-Flash is a third of V4-Pro's price and matches it on simple agent tasks; reasoning_effort="low" cuts output tokens on work that doesn't need deliberation. Both are one-line changes.

The broader judgement: at $0.66 / $1.98 off-peak, V4-Pro remains the cheapest credible frontier-tier option on the market, and the gap to Claude Opus 5 and GPT-5.6 Sol is still large enough that cost-insensitive workloads are the exception rather than the rule. But the sub-dollar era is over, and the sensible planning assumption is that DeepSeek prices capacity, not goodwill. Design so that a 2x move in either direction is a scheduling problem, not a rewrite. If you want the wider context on the model itself — architecture, benchmarks, and how Pro and Flash differ — the DeepSeek V4 complete guide is the pillar for this cluster.

Where to go deeper on DeepSeek V4-Pro

This page is the rate card — the numbers you came for, kept current. Two companion pages cover the parts a price list can't:

• DeepSeek peak hours and off-peak pricing — the timezone conversion table, the weekday-only rule, and how to schedule batch work into the cheap window. Read this if you want to act on the peak/off-peak split rather than just know the rates.
• DeepSeek V4-Pro 0813 benchmarks — every verified SWE-bench Verified and LiveBench score for the GA checkpoint, with sources and methodology caveats. Read this if you are deciding whether the model is good enough, not what it costs.

FAQ

How much does DeepSeek V4-Pro cost?

$0.66 per million input tokens and $1.98 per million output tokens during off-peak hours, doubling to $1.32 and $3.96 during peak. Cache-hit input is $0.022 off-peak and $0.044 at peak. These rates took effect at 16:00 UTC on 16 August 2026 and were verified against DeepSeek's official pricing page on 23 August 2026.

Is DeepSeek V4-Pro cheaper than Claude Opus 5?

Substantially. Claude Opus 5 lists at $5 per million input and $25 per million output. Against V4-Pro's off-peak rates that's 7.6x on input and 12.6x on output. On cached input the gap is widest: Opus 5 charges $0.50 per million cache reads versus V4-Pro's $0.022, a 22.7x difference that dominates long-context bills.

What is DeepSeek's cache-hit pricing?

V4-Pro cache hits cost $0.022 per million tokens off-peak and $0.044 at peak — 30x cheaper than a cache miss. Caching is enabled by default with no code changes. A request only hits when it fully matches a persisted cache prefix unit, so keep reusable content in a stable prefix and check the prompt_cache_hit_tokens field in each response.

When was DeepSeek V4-Pro released?

DeepSeek V4-Pro was released on 24 April 2026 in the DeepSeek-V4 Preview launch, with 1.6T total and 49B active parameters and a 1M-token context window. It reached general availability on 13 August 2026 as checkpoint DeepSeek-V4-Pro-0813, which added agent improvements and a configurable reasoning_effort parameter.

What is DeepSeek V4-Pro's knowledge cutoff?

DeepSeek does not publish an official training-data cutoff for V4. Neither the Hugging Face model card nor the API documentation states one. Third-party model catalogues list April 2026, consistent with the 24 April release date, and that is the best available estimate. DeepSeek did not say whether the August 0813 checkpoint extended it, so test rather than assume.

Did DeepSeek raise its prices in August 2026?

Yes. On 13 August 2026 DeepSeek announced that the flat $0.435 / $0.87 V4-Pro rate would be replaced by peak/off-peak pricing at 16:00 UTC on 16 August. Output rose 2.28x off-peak and 4.55x at peak; cached input rose 12.1x at peak, which produced the widely reported "1,100% increase" headlines. DeepSeek cited capacity allocation as the reason.

When are DeepSeek's peak hours?

01:00–04:00 and 06:00–10:00 UTC, Monday through Friday. All other hours, including the entire weekend, bill at the off-peak rate — about 79% of the week. Those windows correspond to 09:00–12:00 and 14:00–18:00 Beijing time, which is why an entire US business day falls outside them.

What is the cheapest DeepSeek V4 model?

DeepSeek V4-Flash, at $0.22 per million input and $0.66 per million output off-peak — a third of V4-Pro's rate, with a 2,500-request concurrency limit instead of 500. It matches V4-Pro on simple agent tasks and approaches it on reasoning, so it is the right default for bulk extraction, classification, and summarisation.

Do the old deepseek-chat and deepseek-reasoner model names still work?

No. Both aliases were fully retired at 15:59 UTC on 24 July 2026 and now fail. Replace them with deepseek-v4-pro or deepseek-v4-flash explicitly. Some older SDK wrappers and IDE presets still ship the legacy names as defaults, so check pinned dependency versions as well as your own configuration.

Does DeepSeek V4 support image input?

Only through deepseek-v4-flash-vision-exp, an experimental multimodal model released on 21 August 2026. It matches V4-Flash on text and bills images as input tokens at up to 384 tokens each, at V4-Flash rates. V4-Pro itself remains text-only. A free Files API lets you upload an image once and reference it by file_id across requests.