DeepSeek
Cheapest DeepSeek V4 API — 4 Options Compared
Running DeepSeek V4 Flash on 100M input / 25M output tokens a month costs $9 on one platform and $77 on another — an 8x spread on the same model. Here are all four options with real numbers, and why most price quotes you'll find online are measuring the wrong thing.
The short answer
- Cheapest overall: OpenRouter — $9.00/mo for Flash, $115.50 for Pro at the volumes above. It has no affiliate deal with us, which makes it a clean recommendation.
- Cheapest fixed rate: Novita AI — $21.00/mo for Flash, at $0.14/$0.28 around the clock. That is 45% below DeepSeek's own off-peak rate and 73% below peak, with no scheduler to build and no route to re-check. It is still $12/mo more than OpenRouter — you are paying for a rate that does not move.
- On Pro, do not use Novita — it costs roughly double DeepSeek's off-peak rate. Use the official endpoint off-peak, or route through OpenRouter.
- The quotes you're comparing are probably for different builds. Both marketplaces list each model twice, undated and dated, and the two disagree on price and context window. See §4.
1. What each option actually costs
Enter your own volume. Rows are sorted by what you would actually pay, cheapest first.
* OpenRouter rows are the cheapest route for the given model ID, checked 3 Sep 2026. Routes are re-priced continuously — the cheapest host today is not guaranteed to be cheapest tomorrow (see §3).
* DeepSeek Official splits into peak and off-peak: peak is UTC 01:00–04:00 and 06:00–10:00 Monday–Friday, 35 of 168 hours. The other 133 hours, including all weekend, are half price. Check those windows in your timezone →
* Novita quotes a flat rate at every hour. DeepSeek's own endpoint appears as a host inside OpenRouter at exactly its off-peak rate, which is why the Pro rows tie.
2. Flash: where the flat rate wins
On Flash, both resale options beat DeepSeek's own endpoint at every hour of the day:
| Option | Input /1M | Output /1M | Monthly | vs official peak |
|---|---|---|---|---|
| OpenRouter deepseek-v4-flash-0731 Sign up → | $0.05 | $0.16 | $9.00 | −88% |
| Novita AI ref link Sign up → | $0.14 | $0.28 | $21.00 | −73% |
| DeepSeek Official — off-peak Sign up → | $0.22 | $0.66 | $38.50 | −50% |
| DeepSeek Official — peak Sign up → | $0.44 | $1.32 | $77.00 | — |
OpenRouter is the cheapest of the four, and it is also the one we earn nothing from. But it is a routing marketplace: it sends your request to whichever host is cheapest at that moment, and that host changes. Novita charges a flat rate you can put in a budget — at the cost of paying more than the router.
→ Want a flat rate you can forecast? Novita AIref link charges $0.14/$0.28 for Flash around the clock — 45% below DeepSeek's own off-peak rate and 73% below peak, with no scheduler to build.
→ Chasing the lowest possible number? OpenRouter's DeepSeek V4 Flash routes from $0.05/$0.16 — cheaper, but re-check it before you commit spend.
3. Pro: Novita is the wrong answer, and we'll say so
Pro inverts the picture. DeepSeek's off-peak rate on Pro is genuinely competitive, and Novita is not:
| Option | Input /1M | Output /1M | Monthly | Note |
|---|---|---|---|---|
| OpenRouter deepseek-v4-pro-0813 Sign up → | $0.66 | $1.98 | $115.50 | cheapest host is DeepSeek itself |
| DeepSeek Official — off-peak Sign up → | $0.66 | $1.98 | $115.50 | 133 h/week |
| Novita AI ref link Sign up → | $1.60 | $3.20 | $240.00 | 2.1x the cheapest |
| DeepSeek Official — peak Sign up → | $1.32 | $3.96 | $231.00 | 35 h/week |
Nothing here recommends Novita for Pro, so we won't pretend otherwise. If your workload is Pro-heavy, run the official endpoint and shift what you can into off-peak, or route through OpenRouter — where DeepSeek's own host currently lists at $0.66/$1.98, meaning you pay the off-peak rate at 15:00 on a Wednesday.
Novita is also worth a look on Pro if you need dated-build parity — but compare first: its deepseek-v4-pro-0813 entry is $1.32/$3.96, a pass-through of DeepSeek's peak rate. Details in §4.
4. Why the price quote you found is probably wrong
Here is the part that makes cross-platform comparison genuinely hard, and that almost no pricing page mentions: both OpenRouter and Novita list each DeepSeek V4 model twice — an undated entry and a dated build — and the two are priced differently on both platforms, in opposite directions.
On Novita, the dated entries are the expensive ones
| Novita entry | Input /1M | Output /1M | What it actually is |
|---|---|---|---|
deepseek-v4-flash | $0.14 | $0.28 | Novita's own hosting |
deepseek-v4-flash-0731 | $0.44 | $1.32 | Pass-through of DeepSeek's peak rate |
deepseek-v4-pro-0813 | $1.32 | $3.96 | Pass-through of DeepSeek's peak rate |
deepseek-v4-pro | $1.60 | $3.20 | Novita's own hosting |
Those dated figures are recognisable: $0.44/$1.32 is DeepSeek's Flash peak rate and $1.32/$3.96 is its Pro peak rate, to the cent. If you landed on Novita's pricing page, sorted by model name, and saw $0.44 for Flash, you would reasonably conclude Novita is expensive. On Flash it isn't — you were looking at a pass-through resale of DeepSeek's undiscounted peak price. The cheap entry on Novita is the undated one. On Pro the pass-through is $1.32/$3.96, which is cheaper than Novita's own $1.60/$3.20 — and both sit above DeepSeek's off-peak rate (§3).
On OpenRouter, it is the other way round
OpenRouter's dated Flash build is the cheaper of its two Flash entries ($0.05/$0.16 against $0.0679/$0.168) and sees more context — 1,310,720 tokens against 1,048,576. And the undated ID is not "the latest model": OpenRouter's own canonical_slug field shows deepseek-v4-pro is pinned to deepseek-v4-pro-20260423, the April build, three months older than -0813.
This pattern is not a DeepSeek quirk. Across the OpenRouter catalogue, dated entries run from 81% cheaper (qwen3-235b-a22b-2507) to twice the price with a quarter of the context (gpt-3.5-turbo-0613). A comparison quoting a model name without an ID cannot be checked, so treat it as unverified.
5. Flat rate vs routing marketplace
The honest trade-off, since it decides which of the two cheap options suits you:
- OpenRouter is cheaper and less predictable. It is a routing marketplace: many hosts, one price each, and the router picks. We priced
deepseek-v4-pro-0813on 2 Sep 2026 and found its cheapest host at $0.5808/$1.742. Checking again on 3 Sep, that price was gone — the cheapest host is now DeepSeek's own at $0.66/$1.98, and the host that was cheapest had moved to $1.122/$3.366. Twenty-four hours. - Novita is a flat rate. $0.14/$0.28 for Flash, at every hour, on every host. You can forecast it, and you don't have to re-check it monthly.
- DeepSeek direct is cheapest on Pro if your traffic can wait. 79% of the week is already off-peak, so bursty or batch workloads may need no scheduling at all — check your own case.
- Cached input dwarfs all of this. DeepSeek charges $0.007–$0.044/M on cache hits. If your prompts repeat, that matters more than which platform you pick.
6. FAQ
What is the cheapest DeepSeek V4 API?
At 100M input / 25M output tokens a month: OpenRouter, at $9.00 for Flash and $115.50 for Pro. Both are cheapest-route prices checked 3 Sep 2026, and both will move.
Is Novita cheaper than DeepSeek's official API?
For Flash, yes, at every hour: Novita's flat $0.14/$0.28 is 45% below DeepSeek's off-peak $0.22/$0.66 and 73% below peak $0.44/$1.32. For Pro, no — Novita is $1.60/$3.20, roughly double DeepSeek's off-peak $0.66/$1.98. Use the official endpoint or a router for Pro.
Why do I see two different prices for the same DeepSeek model?
Because it is listed twice: an undated entry and a dated build. On Novita the dated entries are pass-throughs of DeepSeek's peak rate and cost more; on OpenRouter the dated Flash build is cheaper and sees more context. They are different builds, not duplicates — see §4.
Does the undated model ID mean "latest version"?
No. On OpenRouter it is pinned to a specific build — deepseek-v4-pro resolves to the 23 April 2026 release. The dated -0813 entry is newer. Check canonical_slug before assuming.
Do I need to schedule around off-peak hours to save money?
Probably not. 79% of the week is already off-peak, weekends included, applied per request with no opt-in. Scheduling only pays if a large share of your traffic lands in the 35 peak hours — and on Flash, alternatives undercut the off-peak rate anyway.
How often should I re-check these prices?
Route prices daily, if the number matters to a decision you're making now. Flat rates like Novita's only need checking when the provider announces a change. Every figure on this page carries a 3 Sep 2026 verified date and is answerable only to that date.
→ Flash workload, and you want a rate that won't move? Start with Novita AIref link — $0.14/$0.28 flat, 45% under DeepSeek's own off-peak rate.
→ Pro workload? Skip Novita and go direct to DeepSeek, or route through OpenRouter at $0.66/$1.98 all day.
Sources
Every figure on this page comes from one of these, fetched and snapshotted 3 Sep 2026:
- DeepSeek Models & Pricing — official peak and off-peak rates, and the peak-window definition.
- OpenRouter model list (JSON) — IDs,
canonical_slug, context lengths. - OpenRouter per-host endpoints (JSON) — cheapest-route prices and host counts.
- Novita AI Pricing — the flat rates quoted above.