What Does DeepSeek Cost for Coding Work in 2026?
DeepSeek-V4-Pro-0813 costs $0.66 per million cache-miss input tokens and $1.98 per million output tokens off-peak, and exactly double both figures during peak hours. Cached input drops to $0.022 per million off-peak, which is where the real money is.
DeepSeek is the only major model vendor that charges by time of day. Two multipliers stack, and together they are the whole story of what DeepSeek actually costs: off-peak halves the bill, and a cache hit is roughly 30x cheaper than a cache miss.
That combination matters most for coding. An agentic coding loop replays the same repository context on every turn, so it is the one workload where high cache hit rates are close to structural rather than lucky.
DeepSeek itself runs no startup credit program. The providers that do are tracked at AI Perks, alongside $7.7M in credits across 194 companies.
Peak and Off-Peak Pricing: The Mechanic Nobody Explains
DeepSeek peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays. Everything else, including all of Saturday and Sunday, is off-peak at half price.
Current list pricing, per million tokens:
| Model | Input, cache hit | Input, cache miss | Output |
|---|---|---|---|
| DeepSeek-V4-Pro-0813, off-peak | $0.022 | $0.66 | $1.98 |
| DeepSeek-V4-Pro-0813, peak | $0.044 | $1.32 | $3.96 |
| DeepSeek-V4.1-Flash, off-peak | $0.003 | $0.15 | not published |
| DeepSeek-V4.1-Flash, peak | $0.006 | $0.30 | not published |
Two things follow from those windows that almost nobody states out loud. Both are derived arithmetic from the published UTC schedule, not quoted figures.
Peak is only about 21% of the week. Seven peak hours per weekday is 35 hours out of 168. Just under 80% of all clock time is already off-peak before you change anything.
A full US working day never touches peak. 9am to 5pm Pacific is 16:00 to 24:00 UTC. 9am to 5pm Eastern is 13:00 to 21:00 UTC. Neither window overlaps a peak block, so US-based teams are paying half price by accident. European mornings are the exception: 08:00 to 10:00 UTC is peak, which is 9am to 11am in London and 10am to noon in Berlin.
There is also a two hour off-peak gap between the two peak blocks, 04:00 to 06:00 UTC, which is a free parking spot for nightly CI jobs and batch refactors.
Why Cache Hits Matter More for Coding Than for Chat
A cache hit on DeepSeek-V4-Pro costs $0.022 per million tokens off-peak against $0.66 for a miss, a 30x gap, and coding agents generate the highest natural hit rates of any workload.
Chat traffic is a stream of unrelated short prompts, so most input is new. A coding agent is the opposite: it carries the same file tree, the same system prompt and the same recent diff through every turn of a long session.
Here is what the two multipliers do to one identical piece of work. Illustrative volumes, arithmetic derived from the list prices above:
| One agent run: 2M input, 200K output | Cost |
|---|---|
| Off-peak, 90% cache hit rate | $0.57 |
| Peak, 90% cache hit rate | $1.14 |
| Off-peak, no caching | $1.72 |
| Peak, no caching | $3.43 |
Same tokens, same model, same output. A 6x spread on identical work, decided entirely by clock time and cache discipline.
Push it to the extreme and the gap widens further: a cached off-peak input token at $0.022 per million against a peak cache-miss token at $1.32 per million is a 60x spread on the same model. No other vendor has a lever like this, which is also why so few tutorials mention it.
At a thousand agent runs a month, the same derived arithmetic gives roughly $570 against roughly $3,430. For a team also paying for Claude or GPT on the hard tasks, AI Perks covers the credit side of that bill.
What Changed in DeepSeek Since V4 Launched
Four releases in five months moved DeepSeek from a flat per-token price to a time-of-day price, and most published guides still quote the old flat numbers.
The sequence, with confidence noted where it should be:
- 24 April 2026: DeepSeek V4-Pro and V4-Flash preview. 1M-token context, up to 384K output tokens, switchable thinking and non-thinking modes. (Confidence: medium.)
- 31 July 2026: DeepSeek-V4-Flash-0731, the stable Flash release. Pricing at the time was $0.14 per million cache-miss input and $0.28 per million output. (Historical context only, not current pricing.)
- 13 August 2026: DeepSeek-V4-Pro-0813 reaches general availability as the flagship reasoning and agent model.
- 16 August 2026: the peak and off-peak split arrives and flat pricing ends.
- Around 10 September 2026: DeepSeek-V4.1-Flash, live in the docs as "DeepSeek-Flash".
If you are reading a DeepSeek cost comparison that shows a single input price and a single output price with no mention of UTC hours, it predates 16 August 2026 and its conclusions are wrong in both directions.
The Endpoint Rename That Breaks Old Code
DeepSeek-V4.1-Flash quietly retired the deepseek-v4-flash and deepseek-v4-flash-vision-exp model names. Code that still passes them will fail.
This is the kind of change that takes a team down at an awkward hour, because the failure surfaces at runtime in whatever service pins the string, not at build time. Older tutorials, Stack Overflow answers and copy-pasted config files all still carry the retired names.
Two practical consequences. First, grep your repository for both strings before your next deploy, not after. Second, treat any DeepSeek tutorial published before September 2026 as unreliable on model identifiers as well as on price.
Does DeepSeek Have a Free Tier or Startup Credits?
The DeepSeek 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 at all.
Be careful with the second and third points, because this is where most sites get DeepSeek wrong. Claims that new API accounts receive free credits are inconsistently reported and are not confirmed on the official pricing page. Anyone quoting a specific new-account credit figure is repeating something DeepSeek has not published.
The absence of a startup program is the more useful fact. Cheap per-token pricing is not the same thing as a grant, and a founder optimising only for list price is leaving the larger number on the table:
| Route | What it actually is |
|---|---|
| DeepSeek chat app | Free, no API access |
| DeepSeek API | Paid from the first token, no stated free tier |
| DeepSeek startup credits | Does not exist |
| Other major model vendors | Many run dedicated startup credit programs |
DeepSeek is the cheapest way to pay. It is not a way to not pay. Those are different strategies and the strongest teams run both: DeepSeek for high-volume cached coding work, granted credits for the frontier models on top. AI Perks tracks which vendors currently fund that second half.
Third-party hosts are a partial workaround. Baseten lists DeepSeek V4.1 Flash input at $0.30 per million and has a standing $30 new-workspace trial credit (confidence: medium, verify before budgeting on it). OpenRouter, Together and Fireworks also host DeepSeek models. Their prices move often enough that the honest advice is to check them yourself on the day rather than trust any number in a blog post, including this one.
How to Cut a DeepSeek Coding Bill
Step 1: Cover the frontier tier with credits. Check which vendors currently fund that tier at getaiperks.com, so DeepSeek only carries the volume work.
Step 2: Check what timezone your agents actually run in. If your CI, nightly refactors or batch evaluation jobs fire inside 01:00 to 04:00 or 06:00 to 10:00 UTC, moving them a few hours halves that line item with no code change.
Step 3: Make the cached prefix stable. Anything that reorders your system prompt, injects a timestamp or shuffles file context turns a $0.022 token into a $0.66 token.
Step 4: Split the models by job. V4-Pro for reasoning and agent loops, Flash for classification, linting and the cheap high-frequency calls around them.
Step 5: Audit your model strings now. The retired Flash endpoint names are a scheduled outage waiting for whoever deploys next.
Frequently Asked Questions
How much does DeepSeek cost for coding?
DeepSeek-V4-Pro-0813 costs $0.66 per million cache-miss input tokens and $1.98 per million output tokens off-peak. Peak hours cost exactly double. Cached input is $0.022 per million off-peak. Real coding bills land between those extremes depending on cache hit rate and clock time.
What are DeepSeek peak and off-peak hours?
Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays. Everything else, including weekends, is off-peak at half price. Derived from those windows, that leaves roughly 80% of the week off-peak, and a standard US working day sits entirely outside both peak windows.
Does DeepSeek offer free API credits?
DeepSeek's chat app is free, but there is no stated API free tier on the official pricing page. Reports of free credits for new API accounts are inconsistent and unconfirmed. DeepSeek also runs no startup credit program, unlike most major vendors 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 names. Any code, tutorial or config still passing those strings will fail at runtime. The rename was not loudly announced, so older samples and copied configs still carry them. Search your repository for both strings and update them to the current Flash identifier.
Is DeepSeek cheaper than Claude or GPT for coding?
On list price, substantially, especially with a high cache hit rate off-peak. On total cost, the comparison changes once startup credits are involved, since DeepSeek grants none and many rivals do. Check what the alternatives will fund before assuming at AI Perks.
Should I use DeepSeek through OpenRouter or Baseten instead?
Third-party hosts add routing flexibility, fallback and sometimes trial credit, but they price independently of DeepSeek's peak and off-peak schedule. Baseten carries a standing $30 new-workspace trial credit. Verify any hosted per-token figure on the day, because these prices change frequently.
DeepSeek is the cheapest way to pay. Credits are the way to not pay.