DeepSeek API pricing in 2026: the budget benchmark
September 14, 2026: the API model name is now deepseek-flash (DeepSeek-V4.1-Flash). The legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted but retired โ requests are served by V4.1-Flash and billed at the Flash price ($0.15 / $0.60 off-peak, down from $0.22 input / $0.66 output). DeepSeek also confirmed V4 Pro stays available after September 14 with unchanged billing.
DeepSeek is the price floor most teams compare everything else against. Its two current models undercut Western flagship APIs by one to two orders of magnitude on list price โ and its cache-hit pricing is the most aggressive in the industry.
| Model | Input | Cached input | Output | Context |
|---|---|---|---|---|
| DeepSeek V4 Pro off-peak | $0.66 | $0.022 | $1.98 | 1M |
| DeepSeek V4 Pro peak | $1.32 | $0.044 | $3.96 | 1M |
| DeepSeek Flash off-peak | $0.15 | $0.003 | $0.60 | 1M |
| DeepSeek Flash peak | $0.30 | $0.006 | $1.20 | 1M |
DeepSeek is the one major vendor that prices by time of day. Peak hours are 01:00โ04:00 and 06:00โ10:00 UTC, MondayโFriday; everything else is off-peak at exactly half the peak rate. In Korea and Japan that peak window lands on 10:00โ13:00 and 15:00โ19:00 local โ the middle of the working day. US and European teams run mostly off-peak by default.
Cache-hit input bills at about 2โ3.3% of the input rate (Flash $0.003, V4 Pro $0.022 off-peak) โ far deeper than the 10โ25% cached rates at OpenAI, Anthropic, Google and xAI. Verified against DeepSeek's official pricing page on August 30, 2026.
How it compares to GPT, Claude and Gemini
| Model | Input | Output | Output vs V4 Pro |
|---|---|---|---|
| DeepSeek V4 Pro (off-peak) | $0.66 | $1.98 | 1ร |
| Gemini 3 Flash | $0.50 | $3.00 | 1.5ร |
| DeepSeek V4 Pro (peak) | $1.32 | $3.96 | 2ร |
| Claude Sonnet 5 | $2.00 | $10.00 | 5.1ร |
| GPT-5.6 Terra | $2.00 | $12.00 | 6.1ร |
| Claude Fable 5 | $10.00 | $50.00 | 25ร |
The gap is still large, but much smaller than it was before DeepSeek's August 16, 2026 price revision, which roughly doubled output rates and introduced peak pricing. Note what the table shows: at peak hours, DeepSeek V4 Pro is more expensive per output token than Gemini 3 Flash. The old shorthand โ "DeepSeek is always an order of magnitude cheaper" โ no longer holds. The honest caveats: there is no batch tier, peak pricing lands on Asian business hours, and for the hardest reasoning tasks the frontier models still lead on quality. For chat, classification, summarization, extraction and most agent plumbing, many teams find V4-class quality more than sufficient.
The caching trick that changes the math
DeepSeek bills repeated prompt prefixes at $0.022 per million tokens off-peak โ about 3.3% of its input rate, where most vendors charge 10โ25%. A chatbot with a long system prompt and an 80% cache-hit rate pays very little for its instruction overhead. Be aware that the August 2026 revision raised the cache-hit rate far more steeply than the headline rate: input went up 52%, cache-hit went up over 500%, so a cache-heavy workload was hit hardest. If your workload is prefix-heavy, DeepSeek's effective price drops even further below the sticker price. Model your exact mix โ cache slider included โ in the calculator, or see how it ranks in the cheapest APIs list.
โ What would YOUR workload cost?Enter your token mix once โ compare all 30 models side by side, with caching and batch discounts.