Is DeepSeek Cheaper Direct or Through a Host?
Direct, in almost every case, and the reason is structural rather than a matter of markup. DeepSeek's low price comes from two multipliers that only exist on the first party API: a half price off-peak window, and a cache hit that costs roughly 30x less than a cache miss on V4-Pro. Resellers bill one flat rate per token, so both multipliers vanish at the door.
That is the whole routing decision in one line. A host is not charging you a modest premium on the same price, it is charging you a flat price against a bill that could have been a fraction of it.
The gap is larger than anyone routing on a leaderboard would guess. Teams chasing this kind of unit economics are usually also chasing free credits, which AI Perks tracks across 194 companies and $7.7M in credits.
The Four Prices DeepSeek Charges for the Same Token
DeepSeek-V4-Pro has four different input prices depending on the hour of day and whether your prompt prefix was cached. No other major vendor prices this way.
Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays. Everything outside those windows, including the entire weekend, is off-peak at half price.
| DeepSeek-V4-Pro-0813 | Off-peak (per 1M) | Peak (per 1M) |
|---|---|---|
| Input, cache hit | $0.022 | $0.044 |
| Input, cache miss | $0.66 | $1.32 |
| Output | $1.98 | $3.96 |
| DeepSeek-V4.1-Flash | Off-peak (per 1M) | Peak (per 1M) |
|---|---|---|
| Input, cache hit | $0.003 | $0.006 |
| Input, cache miss | $0.15 | $0.30 |
The two multipliers stack. A cached off-peak input token at $0.022 against a peak cache miss at $1.32 is a 60x spread on the same model (derived from the table above). On Flash the spread is wider still: $0.003 against $0.30 is 100x, also derived.
Peak hours total 7 hours a day across 5 weekdays, so roughly 79% of the week bills at the off-peak rate by default (derived). You do not have to do anything clever to land there. You have to avoid doing something careless.
What Changed in DeepSeek Pricing in August 2026
The pricing model itself changed on 16 August 2026, when a flat per-token rate gave way to the peak and off-peak split. Any cost comparison written before that date describes a rate card that no longer exists.
The timeline that matters:
- 24 April 2026: V4-Pro and V4-Flash preview, with 1M token context and switchable thinking modes (confidence: medium).
- 31 July 2026: DeepSeek-V4-Flash-0731 ships as the stable Flash release. Last known flat pricing was $0.14 per 1M cache-miss input and $0.28 output (confidence: medium, historical only).
- 13 August 2026: DeepSeek-V4-Pro-0813 reaches general availability.
- Around 10 September 2026: DeepSeek-V4.1-Flash lands, listed in the docs as DeepSeek-Flash.
Set the old $0.14 flat rate beside the current Flash prices and the design of the change becomes obvious. Off-peak cache-miss input at $0.15 sits almost exactly where the old flat rate did, while peak costs $0.30, slightly more than double (derived). DeepSeek did not raise prices so much as install a congestion charge.
What Third Party Hosts Cost and Why They Cannot Match
A host buys capacity and sells tokens, so it has to quote one number that covers its worst hour. That number sits above DeepSeek's off-peak rate and astronomically above its cached off-peak rate.
Two things break in resale:
Time-of-day pricing does not survive. A host serving every timezone cannot pass through a UTC-anchored discount without either arbitrage or an unsellable pricing page.
Cache state is per tenant. Prefix caching pays off when the same system prompt hits the same backend repeatedly. Route that prompt through an aggregator that load balances across providers and you pay cache-miss prices on a prompt you have sent a thousand times.
Verify any host price yourself before committing, because published numbers move and most comparison posts quote rates that expired weeks ago. One figure frames the gap cleanly. Baseten lists DeepSeek V4.1 Flash input at $0.30 per 1M (confidence: medium, verify before quoting), alongside a $30 new-workspace trial credit. The output figure was not reliably confirmed, so it is not quoted here.
| Route | Time-of-day pricing | Prefix cache discount | Verified price signal |
|---|---|---|---|
| DeepSeek first party, V4-Pro | Yes, 2x swing | Yes, 30x | $0.66 in / $1.98 out, off-peak |
| DeepSeek first party, V4.1-Flash | Yes, 2x swing | Yes, 50x (derived) | $0.15 in, off-peak |
| Baseten | Not published | Not published | $0.30 in, V4.1 Flash |
| OpenRouter, Together AI, Fireworks | Not published | Varies by provider | Verify at source |
Against Baseten's listed $0.30, DeepSeek's own off-peak cache-miss input at $0.15 is half the price, and its cached off-peak input at $0.003 is 100x cheaper (derived). That is not a reseller being greedy. It is a flat rate meeting a four-tier one.
OpenRouter, Together AI and Fireworks all host DeepSeek models. Their prices are not quoted here because they were not verified, and a stale price in a cost comparison is worse than no price.
When Paying the Host Markup Is the Right Call
Route through a host when the thing you are buying is not tokens. There are four situations where the premium is rational.
Procurement blocks the China endpoint. If your security review will not approve calling DeepSeek directly, the host is not a markup, it is the only way you ship. This is the single most common reason teams route DeepSeek through a US provider.
You need one invoice and one key across many models. An aggregator that fails over from DeepSeek to a Llama or Qwen endpoint when something is down is selling uptime, not price.
Your traffic is peak-hour and cache-cold anyway. A low volume interactive product serving European mornings sits inside the peak window with short, unique prompts. That workload forfeits both discounts, which narrows the first party advantage considerably.
You are still evaluating. A trial credit such as Baseten's $30 costs nothing and answers the quality question before you wire up a second vendor. Free evaluation credit is the cheapest form of diligence, and AI Perks exists to catalogue where it is available.
The Endpoint Rename That Breaks Routing Configs
DeepSeek-V4.1-Flash quietly retired the deepseek-v4-flash and deepseek-v4-flash-vision-exp model names. Code samples and tutorials written before September 2026 still reference them, and those calls now fail.
This bites hardest on the setup this article is about. A router config holds model identifiers for several providers in one file, most stable for years, so a silently retired name surfaces as an intermittent failure on one route rather than an obvious outage.
Two practical consequences:
- Grep your codebase and your prompt templates for the old names before you debug anything else.
- Host-side model identifiers are namespaced differently from DeepSeek's own, so confirm which exact string your provider exposes rather than assuming the first party name works.
DeepSeek Has No Startup Credit Program. Others Do.
DeepSeek's chat app is free to use, there is no stated API free tier on the official pricing page, and DeepSeek runs no dedicated startup credit program. Claims of API free credits for new accounts circulate widely but are inconsistently reported and not confirmed on the official page. Treat them as unverified.
That is a genuine gap. An API priced this far down is one of the few that will not give a startup a dollar of runway, which makes DeepSeek a cost floor rather than a credits play.
The providers that do run programs are the ones worth applying to, and there are far more of them than most founders know about. AI Perks tracks $7.7M in credits across 194 companies, which is the budget line DeepSeek itself will never cover.
See which providers give credits at getaiperks.com →
Frequently Asked Questions
Is DeepSeek cheaper on OpenRouter or direct?
Direct, structurally. DeepSeek's off-peak half price and its roughly 30x V4-Pro cache-hit discount are first party mechanics that resellers bill as a single flat rate. A host earns its price on compliance, failover and one invoice across models. For raw token cost, first party wins. Credits from providers that run programs are at AI Perks.
When are DeepSeek peak hours?
Peak is 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays. Everything else, including all of Saturday and Sunday, bills at half price. That leaves roughly 79% of the week at the off-peak rate by default (derived), so batch and queued work almost always lands there without any scheduling effort.
How much does a DeepSeek cache hit save?
On DeepSeek-V4-Pro, a cache hit costs $0.022 per million input tokens off-peak against $0.66 for a miss, roughly 30x. On V4.1-Flash the same comparison is $0.003 against $0.15, roughly 50x (derived). Stack the cache discount with off-peak and the spread against a peak miss reaches 60x on V4-Pro.
Why did my deepseek-v4-flash API call stop working?
DeepSeek-V4.1-Flash retired the deepseek-v4-flash and deepseek-v4-flash-vision-exp endpoint names. Any tutorial or code sample written before September 2026 still uses them and will now fail. Update the model identifier, and note that third party hosts namespace their own identifiers differently from DeepSeek's.
Does DeepSeek give free API credits to new accounts?
The DeepSeek chat app is free, but there is no stated API free tier on the official pricing page, and new-account credit claims are inconsistently reported and unconfirmed. DeepSeek also runs no startup credit program. The providers that do are tracked at AI Perks across 194 companies.
What does DeepSeek V4-Pro cost per million tokens?
DeepSeek-V4-Pro-0813 costs $0.66 per million cache-miss input tokens and $1.98 per million output off-peak, doubling to $1.32 and $3.96 during peak hours. Cached input is $0.022 off-peak. That tiering on one model is why a flat reseller rate rarely competes. Offset your other model bills at AI Perks.
DeepSeek sets the price floor. Free credits from everyone else cover the rest.