DeepSeek V4 is a model family introduced with two versions: V4-Pro and V4-Flash. The distinction matters when choosing a model, but the direct DeepSeek API now uses different names for some routes: its documentation recommends deepseek-flash for V4.1-Flash-1017, while listing deepseek-v4-pro separately. A family name, a model version, and a callable API name are not always interchangeable. (DeepSeek V4 Preview Release; DeepSeek Models & Pricing)
What models does DeepSeek V4 include?
DeepSeek’s April 2026 announcement introduced V4-Pro and V4-Flash. It reported a one-million-token context window for both. Its model information lists Pro at 1.6 trillion total parameters, with 49 billion active, and Flash at 284 billion total, with 13 billion active. These figures describe the original V4 models—not necessarily every later version served under a related API name. (DeepSeek V4 Preview Release; DeepSeek-V4-Pro model card)
For the direct DeepSeek API, the current documentation lists:
| API model name | Version listed | Context and output | Vision |
|---|---|---|---|
deepseek-flash |
DeepSeek-V4.1-Flash-1017 | 1M context; up to 384K output | Supported |
deepseek-v4-pro |
DeepSeek-V4-Pro-0813 | 1M context; up to 384K output | Not supported |
DeepSeek says the legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted, but those models are retired: requests using either name are served by DeepSeek-V4.1-Flash-1017 and billed at the Flash price. For new direct API integrations, DeepSeek recommends deepseek-flash. (DeepSeek Models & Pricing; Your First API Call)
A naming caveat for production: DeepSeek’s pricing table lists deepseek-v4-pro as V4-Pro-0813, while a separate official announcement said requests with that name would route to V4.1-Flash starting September 14, 2026, until V4.1-Pro launches. Because these official sources conflict about the Pro route, confirm the model actually served on your account before relying on the alias or estimating its cost. (DeepSeek Models & Pricing; DeepSeek V4.1 Flash announcement)
How should you choose between Pro and Flash?
Start with what your task needs—not with a claim that one version is best for everything.
- Consider Flash when image input is required or when its lower listed direct-API rates suit a high-volume workload. DeepSeek documents image input for
deepseek-flash, but not fordeepseek-v4-pro. (DeepSeek Models & Pricing) - Evaluate Pro when you want to compare its output on a demanding text or coding task and can justify the higher listed rates. Don’t assume it will outperform Flash on your task without testing.
- For either model, test representative prompts and tools, and measure answer quality, errors, latency, and cost on the actual route you plan to use.
As one limited example—not a general ranking—Kilo reported scores of 77/100 for Pro and 60/100 for Flash on a single workflow-backend coding task. Its reviewers found issues in the applications the models generated; those findings are not defects in DeepSeek’s API. (Kilo’s benchmark)
What does direct API pricing look like?
DeepSeek’s current pricing page lists rates per 1 million tokens and displays them with $ price notation. Its off-peak rates are:
| Direct API model | Cache-hit input | Cache-miss input | Output |
|---|---|---|---|
deepseek-flash |
$0.003 | $0.15 | $0.60 |
deepseek-v4-pro |
$0.022 | $0.66 | $1.98 |
Peak rates are twice the off-peak rates. DeepSeek defines peak hours as 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday, excluding Chinese public holidays. All other times—including weekends and Chinese public holidays—are off-peak. Prices may change; these are direct DeepSeek API rates, not prices for another provider or gateway. (DeepSeek Models & Pricing)
Illustrative estimate: At the listed off-peak rates, 1 million cache-miss input tokens plus 100,000 output tokens would cost about $0.21 on Flash ($0.15 + $0.06) or $0.858 on Pro ($0.66 + $0.198). This is a simple token-price calculation, not a measured workload; actual usage and caching can change the bill. Check the live pricing page for applicable rates and schedule.
What should you verify before using a production API?
The direct DeepSeek API documentation lists https://api.deepseek.com as its OpenAI-compatible base URL. That describes DeepSeek’s own API, not another provider’s route. (Your First API Call)
Before moving an application to any production route, confirm:
- The callable model name and the version it serves. A familiar alias does not by itself tell you which underlying version is running.
- The provider and endpoint. Don’t assume an upstream DeepSeek endpoint or behavior applies unchanged through a gateway.
- Required features. Test the input modalities and API tools your application actually uses; DeepSeek’s direct documentation, for example, lists vision for Flash but not Pro.
- The route’s own price and billing rules. Check cache-hit and cache-miss input rates, output rates, currency notation, and any time-based pricing for that provider.
- A representative evaluation. Compare task success, reliability, latency, and total token cost before sending substantial production traffic.
Bottom line
DeepSeek V4 began as a Pro-and-Flash model family, but current direct API names and served versions are not always a one-to-one match with those original family names. The direct pricing table lists Flash as the lower-priced option and supports vision; it lists Pro as text-only, while a separate official announcement creates uncertainty about what the Pro alias currently serves. Confirm the model version and billing for the exact route you intend to use, then test it on your own workload.
For related evaluation considerations, see DeepSeek API: A Practical Evaluation Guide.
Sources
DeepSeek API: Your First API Call
DeepSeek V4.1 Flash announcement
We Tested DeepSeek V4 Pro and Flash Against Claude Opus 4.7 and Kimi K2.6