Short answer: DeepSeek officially announced DeepSeek-V4-Pro general availability on August 13, 2026, for its app, web experience, and API. But GA at launch does not guarantee that the same model is still served today under the same API name. DeepSeek later announced that, starting September 14, requests using deepseek-v4-pro would route to V4.1-Flash until V4.1-Pro launched. Its currently published pricing page lists V4-Pro-0813, without explaining whether that routing change ended. If you need V4-Pro specifically, confirm the live model behind the API ID with DeepSeek before relying on it in production. (DeepSeek GA announcement; routing notice; current pricing and model listing)
Is DeepSeek V4 Pro GA, and where is it available?
Yes—DeepSeek’s official API documentation announced V4-Pro GA on August 13, 2026, saying it was available on the app, web, and API. The announcement also introduced adjustable reasoning effort for V4-Pro and V4-Flash, plus native OpenAI Responses API support. This is stronger evidence than the earlier August 12 social-media report that V4-Pro was “rolling out.” (DeepSeek GA announcement)
There is an important later change for API users. On September 10, DeepSeek said it was phasing out V4-Pro and that, from September 14, 2026 at 04:00 UTC, requests to deepseek-v4-pro would be served by V4.1-Flash at Flash rates, continuing until V4.1-Pro launched. (DeepSeek routing notice)
The current API pricing page lists deepseek-v4-pro with model version DeepSeek-V4-Pro-0813. However, the documentation reviewed here does not say whether the September routing change has ended. So the safe conclusion is: V4-Pro reached GA, but its current API routing is not clear from the published docs alone. Treat the listed model ID as an API entry—not proof that requests currently reach the original V4-Pro model. (DeepSeek models and pricing)
Which model should you consider for your workload?
For the original V4 launch, NVIDIA described V4-Pro as aimed at advanced reasoning, coding, and long-context agents, and V4-Flash as aimed at speed- and efficiency-oriented workloads such as chat, routing, and summarization. NVIDIA listed a 1-million-token context window for both. That is launch-era positioning, not an independent head-to-head test or a guarantee that every provider supports those limits. (NVIDIA Technical Blog)
For the current direct API, DeepSeek’s documentation identifies deepseek-flash as DeepSeek-V4.1-Flash-1017. It also says the legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted, but those models have been retired: requests using those names are served by V4.1-Flash-1017 and billed at the Flash price. DeepSeek’s separate September notice said requests to deepseek-v4-pro would route to V4.1-Flash during the phase-out period. If you require V4-Pro specifically, verify the actual routing first. (DeepSeek models and pricing; routing notice)
These descriptions are selection starting points, not proof of which model will perform best on your application. Test representative prompts and tool workflows, then compare answer quality, latency, and total cost on the provider and model version you plan to use.
What to verify before production API use
Check the documentation for the provider you will actually call. DeepSeek’s current pricing page lists the OpenAI-format base URL as https://api.deepseek.com, the model names deepseek-flash and deepseek-v4-pro, a 1M-token context, and a maximum output of 384K tokens. It also lists an Anthropic-format base URL. Confirm the applicable model behavior and limits for your account and integration before deployment. (DeepSeek models and pricing)
Before shipping, verify:
- Model identity and routing: Does
deepseek-v4-procurrently serve V4-Pro, or another model? - API compatibility: Which endpoint format, authentication, request fields, and response behavior does your client use?
- Limits: What context, output, concurrency, and rate limits apply to your account and selected model?
- Operational terms: Confirm availability and any applicable access conditions directly with the provider.
- Price basis: Check current input, output, and cached-input rates, including any time-based pricing.
DeepSeek’s pricing page lists rates in dollars per 1 million tokens. For deepseek-v4-pro, it lists $1.32 per million uncached input tokens and $3.96 per million output tokens at peak, or $0.66 and $1.98 off-peak. For deepseek-flash, it lists $0.30 input and $1.20 output at peak, or $0.15 and $0.60 off-peak. Cached-input rates are lower; the applicable rate depends on cache status and time. The page defines peak hours as 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday, excluding Chinese public holidays; all other hours, including weekends and Chinese public holidays, are off-peak. Because the V4-Pro routing notice and pricing listing do not fully reconcile, do not use the V4-Pro rate as evidence of which model a request will reach. Prices may change, so check the current pricing page. (DeepSeek models and pricing)
Illustrative cost: At the listed off-peak rates, 1 million uncached input tokens plus 100,000 output tokens would total about $0.86 for V4-Pro ($0.66 + $0.198) or $0.21 for deepseek-flash ($0.15 + $0.06). This is arithmetic from the published rates, not a benchmark; it excludes cache-hit pricing and any changes to rates or routing. (DeepSeek models and pricing)
Finally, keep access routes distinct. DeepSeek API and DeepSeek Chat are DeepSeek services. NVIDIA’s hosted endpoints and NIM/self-hosting are separate NVIDIA routes; their availability does not establish DeepSeek API status or terms. DeepSeek’s announcement confirms the original V4-Pro API GA, while the later routing notice and current model listing leave a specific production question to resolve: what does deepseek-v4-pro serve for your requests now? (DeepSeek GA announcement; NVIDIA Technical Blog)
Sources
DeepSeek-V4.1-Flash: Smarter, Faster, More Efficient
DeepSeek API Docs: Your First API Call
Build with DeepSeek V4 Using NVIDIA Blackwell and GPU-Accelerated Endpoints