DeepSeek Does Not Have One Price Per Token
DeepSeek-V4-Pro has four input prices, not one: $0.022, $0.044, $0.66 and $1.32 per million tokens. Which one you pay depends on the hour of the day and whether the prompt hit cache. That is a 60x spread on the same model, the same prompt, and the same day.
Every "DeepSeek vs ChatGPT" table that prints one number per model is comparing a number that does not exist. DeepSeek is the only major model vendor that charges by time of day. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday, and everything outside those windows, weekends included, bills at exactly half.
Here are the published list prices per million tokens, both halves included.
| Model | Cached input | Fresh input | Output |
|---|---|---|---|
| DeepSeek-V4-Pro, off-peak | $0.022 | $0.66 | $1.98 |
| DeepSeek-V4-Pro, peak | $0.044 | $1.32 | $3.96 |
| DeepSeek-Flash, off-peak | $0.003 | $0.15 | not published |
| DeepSeek-Flash, peak | $0.006 | $0.30 | not published |
That last column is not an omission on our side. DeepSeek publishes cache-hit and cache-miss input rates for the V4.1 Flash line but no output rate, so any guide quoting a Flash output figure is quoting something DeepSeek has not stated. We will not repeat it, and neither should your budget spreadsheet.
The same discipline applies in the other direction. We do not republish OpenAI list prices we have not verified against OpenAI's own pricing page, because those rates move and a stale number in a comparison table is worse than no number. Read this piece for the structure of DeepSeek's bill, then price ChatGPT from the source. AI Perks tracks the third variable, credits, which neither list price includes.
Who Actually Pays DeepSeek's Peak Rate
Peak pricing covers seven hours per weekday, or 35 of the week's 168 hours, so roughly 79% of all clock time bills at half price (derived). Off-peak is not a discount window. It is the normal rate for most of the planet.
Converted to UTC+8, the peak windows run 09:00 to 12:00 and 14:00 to 18:00, with the gap landing across the lunch hour. That is the shape of a working day in DeepSeek's home market. DeepSeek has not stated this as the reason, so treat the fit as an observation rather than a fact, but it predicts exactly who pays more.
| Team location | Working day in UTC | Hours at peak |
|---|---|---|
| San Francisco (UTC-7) | 16:00 to 00:00 | 0 of 8 |
| New York (UTC-4) | 13:00 to 21:00 | 0 of 8 |
| London (UTC+1) | 08:00 to 16:00 | 2 of 8 |
| Berlin (UTC+2) | 07:00 to 15:00 | 3 of 8 |
| Bengaluru (UTC+5:30) | 03:30 to 11:30 | 4.5 of 8 |
| Shenzhen (UTC+8) | 01:00 to 09:00 | 6 of 8 |
A US team running interactive traffic during business hours never touches the peak rate. A European team pays it for the first two or three hours of every morning, which is a concrete argument for scheduling batch and evaluation jobs after lunch rather than before standup.
The Cache Hit Matters More Than the Clock
Moving a workload off-peak halves the bill. Landing it in cache divides it by 30. On DeepSeek-V4-Pro, a cached input token costs $0.022 per million off-peak while a fresh one costs $0.66, exactly 30x more (derived). The off-peak lever is worth 2x. Most teams tune the 2x lever and ignore the 30x one.
Two derived comparisons make the ordering obvious:
- A cached token on the flagship V4-Pro off-peak ($0.022) is 6.8x cheaper than a fresh token on the budget Flash model off-peak ($0.15). Caching beats downgrading.
- Even a cached V4-Pro token at the peak rate ($0.044) is 3.4x cheaper than a fresh Flash token off-peak. Caching beats downgrading and beats waiting for the cheap hours combined.
Raising the cache hit rate from 50% to 90% on V4-Pro cuts blended input cost from $0.341 to $0.0858 per million tokens, a 4.0x reduction (derived) that no amount of scheduling can match.
Three practical consequences:
Prefix stability beats prompt brevity. A long system prompt that never changes is cheaper than a short one you rewrite per request, because only the stable prefix caches.
Put the variable part last. Anything inserted before the retrieved context invalidates the cache for everything after it.
Cache depth is a pricing decision, not a technical one. DeepSeek prices a V4-Pro cache hit at 3.3% of a fresh token and a Flash hit at 2.0% (both derived). Vendors that price cache hits at a shallower fraction of fresh input give you far less to optimise. That ratio, not the headline rate, is what to compare across providers.
What a Real Task Actually Costs
Take a realistic retrieval-augmented agent call: 40,000 input tokens at an 80% cache hit rate, plus 2,000 output tokens. All figures below are derived from the published rates above.
| Scenario | Input cost | Output cost | Total |
|---|---|---|---|
| V4-Pro, off-peak | $0.00598 | $0.00396 | $0.00994 |
| V4-Pro, peak | $0.01197 | $0.00792 | $0.01989 |
| Flash, off-peak | $0.00130 | not published | not quotable |
| Flash, peak | $0.00259 | not published | not quotable |
Two things fall out. First, V4-Pro input on this task costs 4.6x more than Flash input (derived), which is a far smaller gap than the model tiers suggest, because the 80% cache hit is doing most of the work on both. Second, a full task cost for Flash is not computable from published data. If a comparison article gives you one, ask where the output rate came from.
Where the Cheap Reputation Breaks Down
Routing DeepSeek through a third-party host flattens the peak and off-peak split into a single flat rate, so you pay something close to the peak price 100% of the time. Baseten, for example, lists DeepSeek V4.1 Flash input at $0.30 per million, which is identical to DeepSeek's own peak cache-miss rate and double the $0.15 off-peak rate (Baseten's published figure, worth confirming before you budget on it; its output rate was not reliably confirmed). Baseten does carry a standing $30 new-workspace trial credit.
OpenRouter, Together and Fireworks also host DeepSeek models. Their rates are theirs to set and they move, so verify each one yourself rather than trusting a number in a blog post, including this one. The qualitative trade is stable: you gain routing convenience, failover and a single invoice, and you permanently forfeit the 50% off-peak discount. For traffic that runs mostly outside peak, that trade is negative.
What Changed Since the Older DeepSeek Pricing Guides
Every DeepSeek guide written before 16 August 2026 quotes a single flat price. That price no longer exists. If an article gives one number per model with no mention of peak hours, it predates the current structure.
- 24 April 2026: DeepSeek V4-Pro and V4-Flash ship in preview with a 1M token context window, up to 384K output tokens, and switchable thinking and non-thinking modes.
- 31 July 2026: DeepSeek-V4-Flash-0731 becomes the stable Flash release. Pricing at the time was a flat $0.14 per million cache-miss input and $0.28 per million output, with no time-of-day component. That is history, not a rate you can buy today.
- 13 August 2026: DeepSeek-V4-Pro-0813 reaches general availability as the flagship reasoning and agent model.
- 16 August 2026: peak and off-peak pricing arrives.
- Around 10 September 2026: DeepSeek-V4.1-Flash lands, listed in the docs simply as DeepSeek-Flash.
One breaking change deserves its own note, because it silently takes down working code. DeepSeek-V4.1-Flash retired the deepseek-v4-flash and deepseek-v4-flash-vision-exp endpoint names. Tutorials, forum answers and copied code samples still reference those strings and will fail against the live API. Check the current model identifier in DeepSeek's own docs rather than a code sample, since the naming has now changed twice in five months.
DeepSeek Has No Startup Credit Program
The DeepSeek chat app is free. The API is not, and DeepSeek states no free tier on its official pricing page. Claims of free API credits for new DeepSeek accounts circulate widely, are inconsistently reported, and are not confirmed on the official page. Treat them as unverified.
More importantly for founders: DeepSeek runs no dedicated startup credit program. No grant, no stage-based allocation, no subsidised tier. You top up a balance and you spend it. Cheap per token, zero subsidy.
That is the one place the cost advantage inverts. A provider charging several times more per token while granting a young company five figures in credits is cheaper in year one than a cheap provider with no subsidy at all. Cost per token is only the right metric once someone is actually paying cash.
Most of the providers DeepSeek competes with do run credit programs. AI Perks tracks $7.7M in credits across 194 companies, which is the entire layer a list-price comparison misses.
The practical stack: credited capacity from the providers that grant it, DeepSeek-Flash off-peak as the paid default, aggressive prefix caching on anything that runs at peak, and a verified fallback for the hours you cannot shift. Start with the credits, because they are the only part of the bill that can go to zero. See what is currently available.
Frequently Asked Questions
Is DeepSeek cheaper than ChatGPT?
That question has four answers on DeepSeek's side alone. DeepSeek-V4-Pro fresh input runs $0.66 per million off-peak and $1.32 at peak, and a cached token costs $0.022 off-peak. Price your own traffic against OpenAI's current published rates, weighting each DeepSeek rate by the share of calls that actually lands on it. The answer depends on your clock and your cache hit rate, not on the vendor.
What are DeepSeek's peak hours?
Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday. Everything else, weekends included, bills at half the peak rate. That makes roughly 79% of the week off-peak. Teams in US timezones rarely touch peak pricing during normal working hours at all.
Does DeepSeek offer free API credits or a startup program?
No. DeepSeek's chat app is free to use, but no free API tier is stated on its official pricing page, and reports of signup credits are inconsistent and unconfirmed. DeepSeek also runs no startup credit program, unlike many of the providers tracked at AI Perks.
Why did my DeepSeek API call stop working?
DeepSeek-V4.1-Flash retired the deepseek-v4-flash and deepseek-v4-flash-vision-exp model identifiers, quietly and with no deprecation window. Older tutorials and copied samples still reference them and will fail. Pull the current identifier from DeepSeek's live docs before debugging anything else.
What does DeepSeek-Flash cost for output tokens?
DeepSeek has not published this. The V4.1 Flash line lists cache-hit and cache-miss input rates only. Any total-cost-per-task figure you see for Flash is built on an assumed output rate, so treat it as an estimate rather than a price.
Is it cheaper to run DeepSeek through a third-party host?
Usually not, if your traffic runs off-peak. Hosts charge a flat rate with no peak and off-peak split, so you forfeit the 50% off-peak discount permanently in exchange for routing convenience and failover. Verify each host's current rate yourself before switching, since margins vary and change often.
How much do free AI credits actually save a startup?
Enough to reverse the comparison. A provider charging several times more per token but granting five figures in credits is cheaper in year one than a cheap provider with no subsidy. AI Perks tracks $7.7M in credits across 194 companies, which is the layer list-price comparisons ignore entirely.
Subscribe for weekly credit drops
Price per token is the easy number. Price per task, after credits, is the one that shows up on the card.