Short answer: DeepSeek has two text models and two prices for each. deepseek-flash is $0.15 input and $0.60 output per million tokens off-peak, and $0.30/$1.20 at peak; deepseek-v4-pro is $0.66/$1.98 off-peak and $1.32/$3.96 at peak. Cache hits are almost free at $0.003 and $0.022 off-peak. Both models carry a 1M context window and up to 384K output tokens, with no long-context multiplier.
DeepSeek prices per million tokens
| Model | Input, cache hit (peak / off-peak) | Input, cache miss (peak / off-peak) | Output (peak / off-peak) | Context | Max output |
|---|---|---|---|---|---|
deepseek-flash | $0.006 / $0.003 | $0.30 / $0.15 | $1.20 / $0.60 | 1,000,000 | 384K |
deepseek-v4-pro | $0.044 / $0.022 | $1.32 / $0.66 | $3.96 / $1.98 | 1,000,000 | 384K |
USD per 1 million tokens, retrieved 19 November 2026 from DeepSeek's official API pricing page at api-docs.deepseek.com/quick_start/pricing. Peak is listed first, off-peak second, matching the source. The model version behind deepseek-flash is DeepSeek-V4.1-Flash-1017; the legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted but route to the same model and price.
The peak/off-peak clock
DeepSeek defines peak hours as 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday, excluding Chinese public holidays. That is seven hours per weekday. Everything else is off-peak: the other 17 hours of the weekday, all of Saturday and Sunday, and full Chinese public holidays. Off-peak rates are exactly half of peak rates, so the same request costs 50% less outside those two windows.
This is a scheduling problem more than a pricing problem. A batch enrichment job that can run at 11:00 UTC pays half as much as the identical job at 07:00 UTC. A customer-facing assistant that must answer at 02:00 UTC does not get the choice. Before switching providers to chase a headline rate, convert the headline to the hours your traffic actually runs. DeepSeek's off-peak flash rate is among the lowest published, but only for workloads that can move.
Does DeepSeek have a free tier?
No free tier or trial credit is published on DeepSeek's pricing page as of 19 November 2026. The page describes deduction from a topped-up balance or a granted balance, with granted balance used first, but it does not list a free allowance, a daily token cap or a no-card trial. If a third-party page advertises “free DeepSeek tokens,” treat it as a reseller promotion rather than a DeepSeek price.
For a free evaluation path, the practical options are a provider that publishes a free tier, an open-weight self-host, or a small trial from a host that resells DeepSeek. The free LLM API tracker keeps the dated comparison; do not assume the model is free because the weights are downloadable.
What changes the DeepSeek bill
Caching is the biggest lever
DeepSeek's context cache is automatic. When a request shares a prefix with a recent request, the repeated input is billed at the cache-hit rate. On deepseek-flash off-peak that is $0.003 versus $0.15, a 98% discount. On deepseek-v4-pro it is $0.022 versus $0.66, a 96.7% discount. No cache-write or cache-storage fee appears on the pricing page, which makes a stable system prompt, tool schema or reference document unusually valuable.
Cache hits depend on an exact reusable prefix. Put static instructions and documents first and the variable question last. If you interpolate a timestamp or user ID near the top, you break the prefix for every request and pay the miss rate on all of it. Measure the cache-hit token share from the response usage fields before claiming a saving in a budget.
Long context is free; long output is not
DeepSeek does not tier by request size. A 900K-token prompt and a 9K-token prompt pay the same rate per input token, as long as the request fits the 1M window. That is a real advantage over providers that reprice the entire request when it crosses 256K. The catch is output: at $0.60 off-peak for flash, output costs 4× input, and the model supports up to 384K output tokens. A single long generation can cost more than a large retrieval prompt.
Thinking mode is the default on both models. The pricing page publishes one output rate and does not list a separate reasoning-token price, so plan capacity and cost around total output, including reasoning. If you need a cheaper, shorter answer, use non-thinking mode where the task allows it rather than assuming reasoning is free.
Batch pricing is not published
DeepSeek's pricing page does not list a batch endpoint or a batch discount as of 19 November 2026. There is no 50%-off batch line to compare with Groq, Qwen or OpenAI. If a batch path appears in the API docs later, verify whether it reprices output as well as input before assuming a halved bill.
Ancillary APIs
| Item | Unit | Price |
|---|---|---|
| Search API (text) | Per 1,000 successful requests | $0.75 |
| Fetch API | Per 1,000 successful requests | $0.75 |
| File parsing / OCR | Per 1,000 pages | $0.225 |
| Text to speech | Per 10,000 characters | $0.45 list; $0.225 limited-time |
| Speech to text | Per hour of audio | $0.116 list; $0.058 limited-time |
One billing change matters for agent workloads: from 12:00 Beijing time on 18 November 2026, server-side search and web fetch tool calls triggered inside a model call are no longer billed separately. Their results are appended to the model context and billed as model tokens. Standalone Search and Fetch API calls remain $0.75 per 1,000 successful requests.
Worked example: 10M input, 2M output on deepseek-flash
| Scenario | Input cost | Output cost | Total |
|---|---|---|---|
| Off-peak, no cache | 10M × $0.15 = $1.50 | 2M × $0.60 = $1.20 | $2.70 |
| Off-peak, 80% cache hit | 2M × $0.15 + 8M × $0.003 = $0.324 | $1.20 | $1.524 |
| Peak, no cache | 10M × $0.30 = $3.00 | 2M × $1.20 = $2.40 | $5.40 |
| Peak, 80% cache hit | 2M × $0.30 + 8M × $0.006 = $0.648 | $2.40 | $3.048 |
The same workload ranges from $1.52 to $5.40 depending on schedule and cache behaviour, a 3.5× spread with no change to the model or prompt content. That spread is wider than most provider-to-provider gaps at this tier. Model both variables before concluding that DeepSeek is cheaper or more expensive than an alternative.
When a flat plan wins
A flat plan wins when your traffic cannot be scheduled and your prompts cannot be cached. If most of your volume lands inside the two peak windows, or your prompts carry a unique preamble that defeats prefix caching, the effective DeepSeek rate is closer to the peak miss rate, and the headline discount disappears. Add the operational cost of a scheduler and a cache-hit dashboard, and the case for per-token weakens further.
At off-peak flash rates, $60 buys 400 million input tokens or 100 million output tokens. A workload at 300 million input and 40 million output pays $45 + $24 = $69 off-peak, or $138 at peak. A $60 flat plan is cheaper at that volume even before peak amplification, and it removes the incentive to delay work to 22:00 UTC.
Plugsky's Builder plan is $60/month with unlimited fair usage, a 256K input and 32K output ceiling, 10 concurrent requests and 600 RPM; Hobby starts at $8/month. DeepSeek still wins on maximum context (1M versus 256K) and on the absolute off-peak cache-hit rate. Choose DeepSeek for long-document, cache-heavy, batchable work; choose a flat plan for always-on interactive traffic with ordinary prompt sizes.
How to verify these numbers
- Open
api-docs.deepseek.com/quick_start/pricingand confirm both model rows and the peak/off-peak columns. - Convert peak hours to your local time zone; the schedule is published in UTC, not Beijing time.
- Measure cache-hit input share from response usage, not from the size of your system prompt.
- Check the concurrency limit for your model: 2,500 for flash and 500 for v4-pro on the pricing page.
- Re-check the page after any model version change; legacy model names can route to a new version at the same or a different price.
The LLM API cost calculator handles peak share and cache-hit share in one model. Use the rate-limit calculator before assuming the 2,500-concurrency ceiling is the binding limit; token throughput often binds first. The cheapest LLM API guide compares providers by workload, and the model catalogue shows the Plugsky alternatives.
FAQ
How much does DeepSeek charge per million tokens?
On 19 November 2026 deepseek-flash listed $0.30 input on a cache miss, $0.006 input on a cache hit and $1.20 output during peak hours; off-peak is half, at $0.15, $0.003 and $0.60. deepseek-v4-pro listed $1.32, $0.044 and $3.96 at peak, and $0.66, $0.022 and $1.98 off-peak.
When are DeepSeek peak hours?
Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday, excluding Chinese public holidays. Every other hour, plus full weekends and Chinese public holidays, is off-peak and billed at half the peak rate.
Does DeepSeek have a free tier?
No free tier or trial credit is published on DeepSeek's pricing page as of 19 November 2026. The page describes pay-as-you-go billing from a topped-up or granted balance but does not list a free allowance.
Is DeepSeek cheaper with caching?
Yes, dramatically. A cache hit on deepseek-flash costs $0.003 per million input tokens off-peak versus $0.15 on a cache miss, a 98% discount. Caching is automatic and no cache-write fee is published, so a stable prompt prefix is the largest cost lever after off-peak scheduling.
Want one predictable bill across 39 models?
Plugsky is OpenAI-compatible, starts free with no card, and moves to a flat plan when volume is steady. Published RPM, concurrency and daily caps.
Start Free → See plans