DeepSeek V4 Pro Pricing: Full Cost Breakdown (2026)

Budgeting a serious workload on DeepSeek V4 Pro goes sideways the same way every time: the dashboard shows a number that looks fine, and then the invoice is 4x your forecast. You know "Pro is expensive," but nobody tells you what it actually costs per million tokens — or that the input rate you budgeted against is not the input rate you pay. DeepSeek's own pricing page quotes three different input numbers (cache miss, cache hit, and a blended reality between them), the reasoning model writes ~30% more output tokens than a typical model, and the official docs now carry a warning that prices are about to rise "significantly." Each of those alone is survivable. Together, they are how a budget dies.

This guide gives you the DeepSeek V4 Pro pricing picture as of August 13, 2026 — the exact first-party rates, the cache mechanics that decide which number you actually pay, worked cost math for real workloads, and a straight read on the official price-increase notice. All figures come from DeepSeek's official API documentation and HuggingFace model card, cross-checked against Artificial Analysis's independent price/performance data, captured 2026-08-13. DeepSeek has explicitly warned that prices will rise soon, so verify current numbers on the official DeepSeek pricing page before you lock any forecast.

What This Article Solves

You probably know three things already: V4 Pro is DeepSeek's flagship reasoning model, it's big (~1.6T parameters), and it costs more than V4 Flash. What you don't have is a number. This article answers: what does V4 Pro actually cost per million tokens, how does caching change that number, how much does reasoning verbosity inflate the bill, and what does the official "price increase coming" notice mean for your budget. You'll leave with a rule of thumb you can apply to your own volumes in under a minute.

DeepSeek V4 Pro: Official API Pricing

DeepSeek's first-party API (the version that matters for forecasting) lists the current version, DeepSeek-V4-Pro-0813 (released August 13, 2026), at three rates:

Three things matter before you do any math. First, the cache-hit rate is not a rounding error in the docs — at $0.003625/M it is ~99% below the miss rate, and it is the single biggest lever on your bill (more below). Second, the output rate of $0.87/M is strikingly cheap for a reasoning flagship: Artificial Analysis puts the reasoning-class median around $2.20, so V4 Pro outputs run roughly 60% below the competitive set. Third, the concurrency ceiling of 500 requests (a fifth of Flash's 2,500) means at burst you will queue — which matters for batch pipelines and agent farms where throughput, not token price, becomes the bottleneck. If you're comparing tiers, our DeepSeek V4 Flash pricing breakdown covers the cheap end of the family at $0.14/$0.28.

Also worth knowing before you budget: V4 Pro is text-only (no image/audio input), ships thinking mode enabled by default (switchable to non-thinking via the API), and exposes a 1M-token context with up to 384K output tokens. Both API formats are available (https://api.deepseek.com for OpenAI-style requests, /anthropic for Anthropic-format), and it supports JSON output, tool calls, and Responses API.

The Math: Turning Rates into a Rule of Thumb

Per-million-token rates hide the arithmetic that actually hits your wallet. Let's make them concrete. All examples below use the first-party rates above.

• A 1M-token input (cache miss) + 1M-token output job: $0.435 + $0.87 = $1.305. A full million tokens each way, and you're just over a dollar.
• A read-heavy day: 50M input + 5M output = (50 × $0.435) + (5 × $0.87) = $21.75 + $4.35 = $26.10/day — about $783/month with no caching.
• The same day with a hot cache: if 80% of that input hits cache, input becomes (10M × $0.435) + (40M × $0.003625) = $4.35 + $0.145 = $4.50, for a total of $8.85/day — a 66% cut with zero code changes.

Rule of thumb for V4 Pro: count input at ~half a dollar per million (miss) or ~nothing per million (hit), output at ~a dollar per million, and assume a reasoning model writes more output than you planned. A safe planning number for a balanced, cache-mixed workload is the blended rate independent trackers quote — Artificial Analysis computes a 7:2:1 blend (cache-hit input : cache-miss input : output) at ≈ $0.18 per 1M tokens. That is the number to use for rough capacity planning; the scenario table below is for exact budgeting.

Cost Scenarios: Three Real Workloads, Priced

Headline rates become bills in specific shapes. Here are three typical workloads priced end-to-end with first-party rates.

Scenario 3 is the one that changes architecture decisions: a 200K-token knowledge base re-sent with every query is 1 billion input tokens a day — at the cache-miss rate that's ~$439/day, over $13,000/month, for data that doesn't change. With the corpus prefix cached at $0.003625/M, the whole pipeline runs around $344/month. V4 Pro is not viable for this class of workload without caching; with it, the same job is cheap. That's why the cache-hit rate deserves its own section.

One more multiplier the table doesn't show: reasoning verbosity. DeepSeek's flagship is a thinking model by default, and thinking models write more. In Artificial Analysis's independent 9-task evaluation battery, V4 Pro emitted 130M output tokens versus a 100M class median — 30% more output for the same tasks, at $0.87/M that's $113.10 versus $87.00 (the full evaluation cost $135.03 all-in). Budget output tokens at your real volume × 1.3, or switch heavy-but-shallow work to non-thinking mode.

Cache Hit at $0.003625: How KV Caching Cuts Input Cost by 99%

The gap between $0.435 and $0.003625 per million input tokens is the largest pricing spread in DeepSeek's entire rate card, and it's fully mechanical. The API caches the key-value (KV) states of repeated input prefixes — system prompts, tool schemas, document chunks, conversation history — so when a request re-sends tokens the service has seen, it bills them at the near-free cache-hit rate instead of reprocessing them. DeepSeek's official context caching (KV cache) guide documents the mechanism, and the price hit is automatic: you don't opt in, you just get the hit rate when your input prefix matches cache.

Practical implications, in order of impact:

1. Put everything that repeats at the front of the prompt. System prompt, tool definitions, stable context, then the varying part. Cache is prefix-based — the order of your prompt decides your bill.
2. Don't reshuffle prompts between calls. Any edit to a cached prefix invalidates everything after it, sending the whole input back to the $0.435 rate.
3. RAG pipelines must cache the corpus, not re-send it. Scenario 3 above is the difference between a $13K/month and a $344/month line item — and it's the difference between shipping and not shipping.
4. Cache hits are near-free but not free. At $0.003625/M, a gigabyte of cached input costs ~$3.6 — so keep caching, but it's no longer the thing to optimize once hits dominate.

Note that Artificial Analysis lists the cache-hit rate at ~$0.004 (still ~99% off, ranked #8/104 among tracked models); we use DeepSeek's official $0.003625 figure throughout.

The Price-Increase Notice: What It Means for Your Budget

DeepSeek's pricing page currently carries a warning worth quoting plainly: prices will rise significantly across the board in the near future, and the official pricing page is authoritative. There's no date and no magnitude published as of this article's capture. Here's what that does to planning:

• Any forecast built on today's rates is a ceiling, not a baseline. A "significant" increase could plausibly move the cache-hit spread, the headline rates, or both — model your budget with a 1.5x and 2x sensitivity case so the shock is a decision, not a surprise.
• Lock in architecture now, not price later. The part of your bill you control is cache-hit rate, output volume (non-thinking mode, token caps), and provider choice. Those levers work at any price level.
• The window to evaluate alternatives is now. If you're already paying flagship prices for V4 Pro's reasoning, this is exactly the moment to benchmark a comparable flagship on your own prompts — because the "V4 Pro is cheap" positioning erodes the moment the notice lands. We benchmarked our own flagship against this class of model in GLM 5.2 vs DeepSeek V4 Pro; the tier-level math is in the section below.

V4 Pro vs V4 Flash vs GLM 5.2: Picking by Cost-to-Done

Per-token price is half the decision; tokens-to-done is the other half. A cheaper model that needs retries and extra passes on a hard job can outspend a pricier one that finishes once — and on reasoning workloads, output verbosity compounds that gap.

Read the table honestly: V4 Pro's output rate ($0.87) is roughly a third of GLM 5.2's ($4.40), and its input is a third too. On raw per-token economics, V4 Pro is the cheaper flagship — when your workload is dominated by volume and your tasks are V4 Pro-shaped. GLM 5.2's case is the one price-per-token can't show: it targets the same frontier tier with a leaner ~750B/40B-active design and a 1M context, tuned for coding and agentic work, and it's MIT open-source. For an agent farm where a wrong pass costs a re-run, or a coding workflow where you compare cost per completed task rather than per token, the cheaper flagship isn't automatically the cheaper outcome.

The decision framework we actually use:

• Volume and shallow reasoning → V4 Flash. Batch classification, summarization, chat. See the full V4 Flash pricing math.
• Hard reasoning at minimum cost → V4 Pro, with caching engineered in and the price-increase notice priced in.
• Flagship quality with agentic/coding depth as the constraint → benchmark GLM 5.2 against V4 Pro on your real prompts before the notice turns today's rates into a historical footnote. Try GLM 5.2 free on glm5.app — no API key needed — and price it on results, not the rate card; plan details are on the GLM 5.2 pricing page.

FAQ

Will DeepSeek V4 Pro prices increase?

Yes — DeepSeek's official pricing page states that prices will rise significantly in the near future. No date or magnitude was published as of August 13, 2026. Budget a 1.5x–2x sensitivity case and treat today's rates as a ceiling.

How does DeepSeek V4 Pro cache pricing work?

Input tokens that match a cached prefix (system prompt, tool schemas, repeated document chunks) bill at $0.003625/M instead of $0.435/M — roughly a 99% discount, applied automatically via the API's KV cache. Keep repeatable content at the front of the prompt and don't reshuffle prefixes between calls.

How much does DeepSeek V4 Pro cost per token?

Per million tokens: $0.435 input (cache miss), $0.003625 input (cache hit), $0.87 output. On a blended 7:2:1 cache-hit/miss/output ratio, independent trackers put the effective rate at ≈$0.18 per 1M tokens.

Is DeepSeek V4 Pro cheaper than GPT-class or other reasoning flagships?

On output, yes — $0.87/M versus a reasoning-class median of $2.20/M. On cache-miss input ($0.435) it's slightly above the class median ($0.33). Whether it's cheaper end-to-end depends on your cache-hit rate and how much extra output its thinking mode writes (~30% more than the class median in independent testing).

DeepSeek V4 Pro vs V4 Flash — which is better for the price?

V4 Flash ($0.14/$0.28, concurrency 2,500) is dramatically cheaper and built for volume. V4 Pro (roughly 3x on input, ~3x on output, concurrency 500) is the reasoning flagship for the hard 10% of work. Route bulk to Flash, escalate hard reasoning to Pro — and benchmark a flagship alternative like GLM 5.2 before committing to Pro at today's prices.

Bottom Line

DeepSeek V4 Pro pricing as of 2026-08-13: $0.435/M input (cache miss), $0.003625/M input (cache hit, -99%), $0.87/M output, 500 concurrent requests. At the blended 7:2:1 ratio that's an effective ≈$0.18/M — genuinely competitive for a #2-ranked reasoning flagship (II 53, per Artificial Analysis). But three forces can inflate a forecast fast: reasoning verbosity (~+30% output tokens), a 99%-discount cache that only saves you if you engineer for it, and DeepSeek's own warning that prices will rise significantly. Budget with caching built in, model the increase, and — before you commit a multi-month forecast to today's rates — put a comparable flagship through your real workload. Test GLM 5.2 free on glm5.app and see the full plan pricing at glm5.app/pricing to decide which flagship earns your budget once the notice lands.

By the GLM 5 Team. All figures are from DeepSeek's official API documentation and HuggingFace model card, cross-checked against Artificial Analysis as of August 13, 2026; DeepSeek has announced significant price increases are coming, so verify current rates on the official pricing page before production budgeting.

Sources

All facts in this article were captured 2026-08-13 from first-party and independent sources. Prices and versions change often — DeepSeek's own docs warn of a near-term significant price increase — so treat the official pricing page as authoritative before you commit a budget.

• DeepSeek Models & Pricing — Official token pricing ($0.435/$0.003625 input, $0.87 output), model versions, and the official notice of upcoming price increases.
• DeepSeek Context Caching (KV Cache) — Official documentation of the cache-hit pricing mechanism and how cached input prefixes are billed.
• DeepSeek Rate Limits — Official concurrency limits (500 for V4 Pro) for throughput planning.
• DeepSeek-V4-Pro on HuggingFace — Official model card: 1.6T/49B-active MoE architecture, 1M context, MIT license, release details.
• Artificial Analysis — DeepSeek V4 Pro — Independent benchmark (II 53, #2/104), output speed 83.2 tok/s, TTFT 1.63s, verbosity data (130M vs 100M tokens), cache-hit price ($0.004), and blended-rate ($0.18/M) analysis.