DeepSeek API Pricing 2026: Per-Token Rates & Cost Calculator
DeepSeek publishes the cheapest per-token rates of any provider we track, and a cache-hit discount roughly five times deeper than the industry norm. This page shows exactly how far that goes — and the four things you give up.
Prices verified July 28, 2026 against api-docs.deepseek.com/quick_start/pricing.
DeepSeek pricing at a glance
- DeepSeek V4 Flash: $0.14 input / $0.28 output per 1M tokens
- DeepSeek V4 Pro: $0.43 / $0.87 per 1M tokens
- Cache hits: $0.0028 on V4 Flash — a 50x discount where the rest of the industry offers 10x
- No batch discount and no standing free tier
Both models price output at exactly 2x input — a much flatter curve than the 4x–8x you see from OpenAI and the flat 5x from Anthropic. Generation is unusually cheap here relative to prompt processing, which is what makes verbose, self-critiquing and multi-sample architectures affordable on DeepSeek when they are not elsewhere.
DeepSeek API Cost Calculator
Price a DeepSeek workload, then tick on GPT, Claude or Gemini models to see what the identical job costs elsewhere. Note that the batch checkbox does nothing to DeepSeek rates — xAI and DeepSeek publish no batch discount.
Models to compare (5 selected)
| Model | Input cost | Output cost | Per request | Monthly total |
|---|---|---|---|---|
| GPT-5 nanoOpenAICHEAPEST | $38.40 | $12.00 | $0.000025 | $50.40 |
| Gemini 2.5 Flash-LiteGoogle | $76.80 | $12.00 | $0.000044 | $88.80 |
| DeepSeek V4 FlashDeepSeek | $102.14 | $8.40 | $0.000055 | $110.54 |
| DeepSeek V4 ProDeepSeek | $314.94 | $26.10 | $0.00017 | $341.04 |
| Claude Haiku 4.5Anthropic | $768.00 | $150.00 | $0.00046 | $918.00 |
Cheap tokens are wasted if no assistant ever names your tool
You just cut inference costs to the bone. The other half of the equation is demand: our free scan asks ChatGPT about your category five different ways and reports how often your product comes back.
On this page
The 50x cache discount, and why it changes how you write prompts
Almost every provider converged on the same prompt-caching discount: a cache hit costs about 10% of the standard input rate. DeepSeek did not. Here is the same ratio across providers, computed from the verified rates:
| Model | Input / 1M | Cache hit / 1M | Discount |
|---|---|---|---|
| DeepSeek V4 Flash (DeepSeek) | $0.14 | $0.0028 | 50x |
| DeepSeek V4 Pro (DeepSeek) | $0.43 | $0.0036 | 120x |
| GPT-5 (OpenAI) | $1.25 | $0.13 | 10x |
| Claude Sonnet 5 (intro rate) (Anthropic) | $2.00 | $0.20 | 10x |
| Gemini 2.5 Flash (Google) | $0.30 | $0.03 | 10x |
| Grok 4.3 (xAI) | $1.25 | $0.20 | 6x |
DeepSeek V4 Pro reaches a 120x discount. At $0.0036 per million tokens, a cached prefix is not cheap — it is numerically irrelevant. You could hold a 100,000-token reference document in the cached prefix of every single request and the cached portion would contribute roughly $0.00036 per call.
This inverts the usual prompt-engineering advice. On OpenAI or Anthropic you are always trading prompt richness against cost, and trimming context is a real lever. On DeepSeek, a stable prefix is effectively free once cached, so the optimisation is not "make the prompt shorter", it is "make more of the prompt stable". Move everything invariant — instructions, schemas, few-shot examples, reference material — to the front, and keep the variable tail as small as you can. Length stops mattering; volatility starts mattering instead.
One caveat worth stating plainly: caching only helps if requests actually share a prefix and arrive close enough together to hit a warm cache. A pipeline where every request carries a different document has nothing to cache, and on that shape DeepSeek's advantage collapses back to its ordinary — still very low — miss rate.
DeepSeek API price table, per 1M tokens
| Model | Input / 1M | Cache hit / 1M | Output / 1M | Output ÷ input | Batch |
|---|---|---|---|---|---|
DeepSeek V4 Flashdeepseek-v4-flash | $0.14 | $0.0028 | $0.28 | 2.0x | Not published |
DeepSeek V4 Prodeepseek-v4-pro | $0.43 | $0.0036 | $0.87 | 2.0x | Not published |
V4 Pro costs about 3.1x V4 Flash on both directions — an unusually tidy ratio, and a rare case where the "bigger model" premium is the same on input and output. Even V4 Pro at $0.87 output is cheaper than every non-DeepSeek model on this site except the Gemini and GPT lite tiers.
One email, only when a rate on this page moves
DeepSeek's advantage is a moving target — the frontier labs keep cutting toward it. We re-check all five providers and write to you when the gap changes.
Already on the calendar: Claude Sonnet 5's introductory $2 / $10 per 1M tokens expires Aug 31, 2026 and goes to $3 / $15 — a 50% increase on the same workload. Prices here were last re-verified July 28, 2026.
What the price gap actually buys: a bulk classification job
Two million records a month, each a 600-token prompt returning a 15-token label, 40% of input served from cache. This is the workload shape where per-token price dominates everything else — tiny requests, enormous volume, no reasoning, no long context.
| Model | Input | Output | Monthly total | vs V4 Flash |
|---|---|---|---|---|
| GPT-5 nano (OpenAI) | $38.40 | $12.00 | $50.40 | 0.5x |
| DeepSeek V4 Flash (DeepSeek) | $102.14 | $8.40 | $110.54 | 1.0x |
| Gemini 2.5 Flash (Google) | $230.40 | $75.00 | $305.40 | 2.8x |
| DeepSeek V4 Pro (DeepSeek) | $314.94 | $26.10 | $341.04 | 3.1x |
| GPT-5 (OpenAI) | $960.00 | $300.00 | $1,260 | 11.4x |
| Claude Sonnet 5 (intro rate) (Anthropic) | $1,536 | $300.00 | $1,836 | 16.6x |
At two million requests a month the same job spans from $50.40 to $1,836. That is the difference between a line item nobody notices and one that needs a budget conversation — for identical output on a task where frontier reasoning contributes nothing.
Do not over-read this. Classification is the single most flattering workload for a cheap model, because quality differences barely show up on a well-specified 12-label task. Run the same comparison on the calculator with a coding-agent shape — 30,000 input, 4,000 output — and the absolute gap grows but the decision gets much harder, because a model that produces a wrong diff costs you engineering time that dwarfs every figure in this table.
The missing batch discount narrows the gap more than you would expect
OpenAI, Anthropic and Google all take 50% off both directions for asynchronous work. DeepSeek does not publish a batch discount at all. For an offline pipeline that is a real handicap, and comparing DeepSeek's standard rate against a competitor's standard rate overstates the saving by 2x.
The same two-million-record job, with the batch discount applied wherever the provider offers one:
| Model | Batch applied? | Monthly total |
|---|---|---|
| GPT-5 nano (OpenAI) | Yes — 50% off | $25.20 |
| DeepSeek V4 Flash (DeepSeek) | No batch discount published | $110.54 |
| Gemini 2.5 Flash (Google) | Yes — 50% off | $152.70 |
| Claude Sonnet 5 (intro rate) (Anthropic) | Yes — 50% off | $918.00 |
The ordering does not necessarily change, but the multiples do — sometimes dramatically. The rule to take away: if your workload is fully offline, benchmark DeepSeek's standard rate against everyone else's batch rate. If your workload is interactive, benchmark standard against standard, and DeepSeek's advantage stands at full size.
Four things you give up, none of them about model quality
1. Data jurisdiction
DeepSeek is a Chinese company. If your data handling is governed by customer contracts, a procurement review, a SOC 2 scope or any regulatory regime, where inference happens is the first question your legal and security functions will ask — and at a large share of companies it ends the conversation before price is even discussed. Establish that answer before you spend a sprint on migration.
2. No batch discount
Covered above. It removes the single biggest lever available on offline workloads and halves DeepSeek's effective advantage on that shape.
3. No standing free tier
You pay from the first token. The rates are low enough that this rarely blocks a prototype, but it does mean you cannot hand an unfunded experiment to a team the way you can with the free Gemini Flash tier.
4. Ecosystem depth
Fewer first-class framework integrations, fewer battle-tested SDK paths, and a much smaller pool of people who have already hit your specific production bug. This is a real engineering-time cost that does not appear on any pricing page, and on a small team it can exceed the token saving.
DeepSeek API pricing FAQ
How much does the DeepSeek API cost?
As of July 28, 2026, DeepSeek V4 Flash is $0.14 per million input tokens and $0.28 per million output tokens, and DeepSeek V4 Pro is $0.43 / $0.87. Cache hits are billed at $0.0028 and $0.0036 respectively. These are the lowest published per-token rates of any provider on this site — V4 Flash input is about 9x cheaper than GPT-5 and its output is roughly 36x cheaper.
Is DeepSeek's cache-hit discount really 50x?
Yes, and it is the most unusual number in LLM pricing right now. OpenAI, Anthropic and Google all price a cache hit at roughly 10% of the standard input rate — a 10x discount. DeepSeek V4 Flash charges $0.14 for a miss and $0.0028 for a hit, a 50x discount, and V4 Pro is deeper still at 120x. In practice this means cached input on DeepSeek is close enough to free that it drops out of your cost model entirely, which changes how you architect prompts: on DeepSeek you have no reason to trim a stable prefix, only a volatile one.
Does DeepSeek have a batch API discount?
No. DeepSeek does not publish a 50% asynchronous batch discount, and neither does xAI. OpenAI, Anthropic and Google all do. This matters more than it sounds for offline workloads: a batched GPT-5 mini job effectively costs $0.125 per million input tokens, which narrows the gap against DeepSeek's headline rate considerably. If your workload is entirely offline — nightly summarisation, backfills, eval runs — you should compare DeepSeek's standard rate against competitors' batch rates, not their standard rates, or you will overstate the saving.
Does DeepSeek have a free tier?
No standing free tier. New accounts may receive promotional credit, but there is no perpetual free allowance of the kind Google publishes for several Gemini Flash models. If free-to-build is the requirement, Gemini is the API to look at; if cheapest-at-scale is the requirement, DeepSeek is.
Why is DeepSeek so much cheaper than OpenAI or Anthropic?
A combination of things, and honestly none of them are visible from a pricing page. Architecture is part of it — DeepSeek's models use a mixture-of-experts design that activates only a fraction of total parameters per token, so the serving cost per token is genuinely lower than a comparable dense model. Beyond that, pricing is a strategic choice as much as a cost-recovery calculation, and a challenger pricing aggressively to win developer share is a familiar pattern. The practical implication for you is the same either way: a price set strategically can be changed strategically, so do not architect a product whose unit economics only work at today's DeepSeek rate.
What do I give up by moving to DeepSeek?
Four concrete things, none of which are about the model output. No batch discount, so your offline pipeline cannot be halved. No standing free tier for prototyping. A smaller ecosystem — fewer SDK integrations, fewer framework adapters, fewer people who have already hit and solved your specific bug. And data jurisdiction: DeepSeek is a Chinese company, and if your data handling is governed by customer contracts, procurement review or a regulatory regime, where inference happens is a question your legal team will ask before your engineering team ships. That last one kills the deal at a lot of companies regardless of price.
Is DeepSeek cheaper than the small models from OpenAI and Google?
Less dramatically than the headline comparison suggests. Against GPT-5 or Claude Opus the gap is 10x to 40x. Against the deliberately cheap tiers it narrows sharply: GPT-5 nano is $0.05 / $0.40 — actually cheaper than DeepSeek V4 Flash on input — and Gemini 2.5 Flash-Lite is $0.10 / $0.40. DeepSeek V4 Flash still wins on output against both, and wins enormously on cached input, but "DeepSeek is 30x cheaper" is a claim about frontier models, not about the small-model tier where most high-volume production traffic actually runs.
Should I self-host DeepSeek instead of using the API?
Almost certainly not for cost reasons. At these rates the arithmetic is brutal: a single always-on high-memory GPU instance costs more per month than tens of millions of tokens through the API, and you also inherit capacity planning, model updates, batching, and an on-call rotation. Self-hosting an open-weights model is a decision about data residency, air-gapped deployment or customisation. It is not a cost optimisation at DeepSeek's API prices, and teams that frame it as one are usually comparing GPU rental against a frontier-model bill rather than against DeepSeek's own.
Do DeepSeek's prices include reasoning tokens?
Reasoning output is billed at the output rate, as it is on every provider. The difference is that it hurts far less here: at $0.28 per million, several thousand hidden reasoning tokens per request costs a fraction of a cent. On a frontier model at $25 to $180 per million output, the same reasoning burn is the dominant line on the invoice. Cheap output is what makes reasoning-heavy architectures — self-critique loops, multi-sample voting, verbose chain of thought — economically viable at all.
Per-token pricing for the other providers
Every page in this set reads from the same verified price table, so the numbers never disagree with each other.
OpenAI API pricing
GPT-5.6, GPT-5, mini and nano rates, plus the cost calculator.
Claude API pricing
Anthropic's tier ladder, cache-write premiums and the Sonnet 5 intro rate.
Gemini API pricing
Google's free tier, the 200k context price cliff and the Batch/Priority service tiers.
Grok API pricing
xAI's unusually flat input/output gap and what it means for output-heavy apps.
LLM API pricing comparison
All five providers head to head, with a recommendation per workload shape.
Related
ChatGPT already recommends your DeepSeek-powered app. Does it recommend yours?
If you're building in LLM APIs, run a free AI-visibility scan on your own product — we ask ChatGPT across 5 prompt angles and score how often you get named. ~30 seconds, no signup, no card.
📬 Get the best new AI tools delivered weekly
One concise email with fresh launches, trending picks, and featured standouts.