DeepSeek V4
DeepSeek V4 is DeepSeek's current model family, succeeding V3 and R1, and includes V4.1-Flash as the default workhorse model and V4-Pro for frontier-tier reasoning, agentic and coding work. Pricing is per million tokens, with V4.1-Flash starting at $0.15 per million input tokens off-peak.
Facts checked on against the sources listed at the end of this page.
- Pricing
- Open sourcePaid from $0.15 per million input tokens off-peak (V4.1-Flash, cache miss)
- Open source
- YesMIT (weights and code)
A good fit for
- Developers wanting a low-cost, high-performing open-weight model family via API or self-hosting
- Cost-sensitive projects that can take advantage of off-peak and cache-hit pricing discounts
- Frontier-tier reasoning, agentic and coding work using V4-Pro
- Teams wanting the flexibility to self-host open weights instead of relying solely on an API
Not a good fit for
- Users still relying on legacy V3/R1-era API model names, which have been retired in favor of V4
- Teams needing guaranteed flat pricing without off-peak/peak variation
- Applications requiring a vendor-hosted proprietary model rather than an open-weight one
DeepSeek V4 pricing
V4.1-Flash costs $0.15 per million input tokens off-peak ($0.30 peak) with cache-hit rates as low as $0.003/M, and $0.60 per million output tokens off-peak ($1.20 peak). V4-Pro costs $0.66 per million input tokens off-peak ($1.32 peak) with cache-hit rates from $0.022/M, and $1.98 per million output tokens off-peak ($3.96 peak). Weights are free to download and self-host under the MIT license.
| Plan | Price | What you get |
|---|---|---|
| V4.1-Flash | $0.15/M input (off-peak), $0.60/M output (off-peak) | Default workhorse model for everyday use, with peak-hour and cache-hit pricing variations. |
| V4-Pro | $0.66/M input (off-peak), $1.98/M output (off-peak) | Frontier-tier model for the hardest reasoning, agentic and coding work. |
Prices in USD, checked on 27 September 2026. Current prices on deepseek.ai.
Main features
- V4.1-Flash
- The default workhorse model, balancing cost and performance for everyday use.
- V4-Pro
- A frontier-tier model for the hardest reasoning, agentic and coding tasks.
- Open weights
- Released under the MIT license, allowing self-hosting and modification.
- Off-peak/peak and cache pricing
- API pricing varies by time of day and offers discounts for cached input tokens.
About DeepSeek V4
DeepSeek V4 is the current generation of DeepSeek's model family, replacing the earlier V3 and R1 models, and is offered in two main variants: V4.1-Flash, described as the default workhorse model, and V4-Pro, positioned for frontier-tier reasoning, agentic and coding work.
As with earlier DeepSeek models, V4 is released as open weights under the MIT license, so it can be downloaded and self-hosted in addition to being accessed through DeepSeek's own API and third-party inference providers.
DeepSeek's API pricing includes off-peak and peak rates, along with cache-hit discounts for repeated context, and DeepSeek retired the legacy V3/R1-era API model names in July 2026, routing that traffic to the V4 family.
- Made by
- DeepSeek
- Free plan
- Yes, open-weight models free to download and self-host; API access is pay-as-you-go
- Available on
- API, Self-hosted (open weights), Various inference providers
DeepSeek V4 alternatives
Questions about DeepSeek V4
What is the difference between V4.1-Flash and V4-Pro?
V4.1-Flash is the default, cost-efficient workhorse model, while V4-Pro is positioned for the hardest reasoning, agentic and coding tasks.
Is DeepSeek V4 open source?
Yes, the model weights are released under the MIT license and can be self-hosted; API access is also available on a pay-as-you-go basis.
Does DeepSeek V4 pricing vary by time of day?
Yes, off-peak rates are lower than peak rates, and cached input tokens receive a significant discount.
What happened to DeepSeek V3 and R1?
DeepSeek retired their legacy API model names on July 24, 2026, routing that traffic to the V4 family.
Sources and links
Work on DeepSeek V4? Claim this listing. Something wrong or out of date? Suggest a correction. Stats come from the public GitHub and npm APIs and update every day.