Last verified: October 7, 2026
DeepSeek V4 Pro is the Pro route of DeepSeek’s V4 model family. In the current DeepSeek API model table, the API model name is deepseek-v4-pro, and the listed model version is DeepSeek-V4-Pro-0813. It supports a 1M-token context window, a 384K maximum output, Thinking Mode, JSON Output, Tool Calls, Responses API, Anthropic API, Chat Prefix Completion, and FIM in non-thinking mode. It does not support Vision in the current model table. DeepSeek Models & Pricing
The important nuance is that deepseek-v4-pro has had a confusing lifecycle. DeepSeek’s September 10, 2026 V4.1 Flash announcement said DeepSeek was phasing out V4 Pro and that deepseek-v4-pro requests would route to V4.1 Flash after September 14, 2026 until V4.1-Pro launched. But DeepSeek’s API Change Log later says the company decided to continue providing API service for DeepSeek V4 Pro after September 14, 2026, with billing unchanged. The current Models & Pricing page also still lists deepseek-v4-pro separately from deepseek-flash. DeepSeek V4.1 Flash announcement DeepSeek API Change Log DeepSeek Models & Pricing
The practical answer: do not treat DeepSeek V4 Pro as dead, but do not treat it as the default choice for every new project either. For most new cost-sensitive API work, start by evaluating deepseek-flash. Use deepseek-v4-pro only if your own tests show that its Pro behavior is worth the higher price, lower concurrency limit, and lack of Vision support.
Quick answer: should you use deepseek-v4-pro?
Use deepseek-v4-pro only when your workload needs the Pro route and your evaluation shows better results than deepseek-flash. The current DeepSeek model table still lists deepseek-v4-pro, but it is much more expensive than deepseek-flash, has a lower concurrency limit, and does not support Vision. DeepSeek Models & Pricing
| Question | Practical answer |
|---|---|
| What is the current API model name? | deepseek-v4-pro. |
| What model version does it point to in the current table? | DeepSeek-V4-Pro-0813. |
| Is V4 Pro definitely retired? | No. The current model table still lists it, and the Change Log says service continues after September 14, 2026. |
| Did DeepSeek announce a V4 Pro phaseout? | Yes. The V4.1 Flash announcement said V4 Pro would be phased out/routed, but the later/current docs must be checked before production decisions. |
| Does V4 Pro support Vision? | No, not in the current Models & Pricing table. |
Is it cheaper than deepseek-flash? | No. It is substantially more expensive in the current pricing table. |
| Should new projects start with it? | Usually no. Start with deepseek-flash unless you specifically need to test Pro behavior. |
| Can it be self-hosted? | The official Hugging Face model card lists DeepSeek-V4-Pro-0813 weights under MIT, but self-hosting a large model is not cheap or simple. |
What is DeepSeek V4 Pro?
DeepSeek V4 Pro is the higher-cost Pro route in the DeepSeek V4 family. The official Hugging Face model card identifies DeepSeek-V4-Pro-0813 as the official release of DeepSeek-V4-Pro, superseding the preview version. It says the model was improved for agentic capabilities and production environments, and that it is built on the DeepSeek-V4-Pro Preview structure with a DSpark speculative decoding module attached. DeepSeek-V4-Pro-0813 model card
In plain terms, V4 Pro is not the same thing as DeepSeek R1, deepseek-reasoner, V4 Flash, V4.1 Flash, or deepseek-flash. It is a specific Pro model route, and the current official API model ID to check is deepseek-v4-pro.
For hosted API users, the most important source is the current DeepSeek Models & Pricing page. For open-weight or local deployment users, the most important source is the DeepSeek-V4-Pro-0813 model card. Do not assume that behavior from one layer automatically applies to the other.
Current model status: V4 Pro, V4.1 Flash, V4 Flash, and old aliases
The model names around DeepSeek V4 can be confusing because several old names still appear in older tutorials. Use this table to separate the current API model names from old or legacy names.
| Item | What it is | Current official status | Should you use it today? | Practical note |
|---|---|---|---|---|
| DeepSeek V4 Pro | Pro route of the V4 family. | Listed in the current Models & Pricing page as DeepSeek-V4-Pro-0813. | Only if testing shows it is worth the higher cost. | Use deepseek-v4-pro if the current docs still list it and your workload justifies it. |
| DeepSeek-V4-Pro-0813 | Official GA release of V4 Pro. | Official Hugging Face model card says it supersedes the preview version. | Yes, if you specifically need V4 Pro weights or behavior. | Model card lists MIT license and local deployment paths. |
deepseek-v4-pro | Hosted API model ID for V4 Pro. | Current model table still lists it separately from deepseek-flash. | Use carefully after checking current docs. | Not a good default for image input or lowest-cost workloads. |
| DeepSeek V4 Pro Preview | Earlier preview version. | Superseded by DeepSeek-V4-Pro-0813 according to the model card. | No, not as the current Pro reference. | Old benchmark/tutorial content may refer to the preview and be outdated. |
| DeepSeek V4.1 Flash | Current Flash route. | Listed as deepseek-flash in the current model table. | Usually yes, for new cost-sensitive API work. | Supports Vision and is much cheaper than V4 Pro. |
deepseek-flash | API model ID for V4.1 Flash. | Current preferred Flash model name. | Yes, for most new DeepSeek API projects. | Best default starting point unless Pro behavior is specifically needed. |
deepseek-v4-flash | Legacy V4 Flash API name. | Accepted for compatibility, but its old model is retired and requests are served by V4.1 Flash. | Avoid in new code. | Use deepseek-flash instead. |
deepseek-v4-flash-vision-exp | Legacy experimental Vision route. | Accepted for compatibility, but the old model is retired and requests are served by V4.1 Flash. | Avoid in new code. | Use deepseek-flash for Vision. |
| DeepSeek R1 | Historical/open-weight reasoning model family. | Still important as open weights and historical reasoning work, but not the V4 Pro API route. | Use only when you specifically need R1. | Do not treat R1 tutorials as current V4 Pro guidance. |
deepseek-chat | Old general chat alias. | DeepSeek’s V4 changelog said legacy API names would be discontinued after July 24, 2026. | No, not as a new default. | Use current model IDs from the model table. |
deepseek-reasoner | Old reasoning alias tied to R1-era usage. | Legacy/old API guidance, not the current V4 Pro default. | No, not as a new default. | Use current Thinking Mode and current model IDs instead. |
What is DeepSeek-V4-Pro-0813?
DeepSeek-V4-Pro-0813 is the official release of DeepSeek V4 Pro that replaced the preview version. The model card says it has greatly enhanced agentic capabilities and production-environment performance compared with the preview version. It also says DeepSeek-V4-Pro-0813 is built on the V4-Pro Preview model structure with a DSpark speculative decoding module attached. DeepSeek-V4-Pro-0813 model card
The current DeepSeek Models & Pricing page maps the hosted API model ID deepseek-v4-pro to the model version DeepSeek-V4-Pro-0813. That means when developers talk about DeepSeek V4 Pro today, they should usually mean deepseek-v4-pro / DeepSeek-V4-Pro-0813, not the older preview version. DeepSeek Models & Pricing
What happened on August 13, 2026?
DeepSeek’s API Change Log says the GA release of DeepSeek-V4-Pro was rolled out on the app, web, and API on August 13, 2026. It says the API calling method remained unchanged: set the model name to deepseek-v4-pro to use the latest version. The same entry describes improved agent capabilities and native Responses API support. DeepSeek API Change Log
The Hugging Face model card reinforces this lifecycle: DeepSeek-V4-Pro-0813 is the official release that superseded V4 Pro Preview. This distinction matters because benchmark claims, pricing assumptions, and examples from the preview period may not describe the current V4 Pro route. DeepSeek-V4-Pro-0813 model card
What happened after V4.1 Flash launched?
On September 10, 2026, DeepSeek introduced DeepSeek-V4.1-Flash. The official announcement said V4.1 Flash was live on the API with native multimodal support and told developers to set the model to deepseek-flash. It also said V4 Flash and V4 Flash Vision Exp were retired, with their old names temporarily routed to V4.1 Flash. DeepSeek V4.1 Flash announcement
The same announcement said DeepSeek was phasing out V4 Pro and that, starting September 14, 2026, deepseek-v4-pro requests would route to V4.1 Flash at V4.1 Flash rates until V4.1-Pro launched. That created understandable confusion for developers using V4 Pro. DeepSeek V4.1 Flash announcement
However, DeepSeek’s API Change Log now says the company decided to continue providing API service for DeepSeek V4 Pro after September 14, 2026, with billing unchanged. The current Models & Pricing page also lists deepseek-v4-pro separately. The safest guidance is to treat V4 Pro as an available but specialized route only when the live official documentation still lists it. DeepSeek API Change Log DeepSeek Models & Pricing
Does deepseek-v4-pro still work?
Based on the current official model table, deepseek-v4-pro is still listed as a supported API model ID. Based on the current API Change Log, DeepSeek decided to continue providing V4 Pro API service after September 14, 2026. But because the earlier V4.1 Flash announcement described a phaseout/routing plan, developers should verify the live Models & Pricing page and Change Log before production deployments. DeepSeek Models & Pricing DeepSeek API Change Log
In other words, do not say “V4 Pro is dead,” and do not say “V4 Pro always routes to Flash.” The current official status is more nuanced: V4 Pro remains listed, but DeepSeek’s model lifecycle around V4.1 Flash means developers should check current documentation before relying on it.
API features and supported capabilities
The current DeepSeek model table lists these capabilities for deepseek-v4-pro:
| Capability | deepseek-v4-pro support | Practical note |
|---|---|---|
| Model version | DeepSeek-V4-Pro-0813. | Current listed Pro model version. |
| OpenAI-format base URL | https://api.deepseek.com. | Use with OpenAI-compatible clients. |
| Anthropic-format base URL | https://api.deepseek.com/anthropic. | Use with Anthropic-compatible clients. |
| Thinking Mode | Supported. | Supports both thinking and non-thinking modes, with thinking as default in the model table. |
| Context length | 1M. | Useful for long-context tasks, subject to quality and cost testing. |
| Max output | 384K maximum in the current model table. | Long output can become expensive and slow; plan carefully. |
| JSON Output | Supported. | Useful for structured extraction and machine-readable responses. |
| Tool Calls | Supported. | Useful for agents and function-calling workflows. |
| Responses API | Supported. | Useful for Responses-style workflows. |
| Anthropic API | Supported. | Useful for Anthropic-compatible tooling. |
| Chat Prefix Completion | Supported, listed as Beta. | Use only when that completion pattern fits your application. |
| FIM Completion | Supported in non-thinking mode only. | Useful for code-completion-style workflows. |
| Vision | Not supported. | Use deepseek-flash if you need image input through the official API. |
Pricing and concurrency
DeepSeek prices API usage per 1M tokens. The current table separates cache-hit input, cache-miss input, and output tokens, and it uses peak and off-peak rates. deepseek-v4-pro is much more expensive than deepseek-flash. DeepSeek Models & Pricing
| Model | Input cache hit | Input cache miss | Output | Concurrency limit |
|---|---|---|---|---|
deepseek-v4-pro off-peak | $0.022 / 1M tokens | $0.66 / 1M tokens | $1.98 / 1M tokens | 500 |
deepseek-v4-pro peak | $0.044 / 1M tokens | $1.32 / 1M tokens | $3.96 / 1M tokens | 500 |
deepseek-flash off-peak | $0.003 / 1M tokens | $0.15 / 1M tokens | $0.60 / 1M tokens | 2500 |
deepseek-flash peak | $0.006 / 1M tokens | $0.30 / 1M tokens | $1.20 / 1M tokens | 2500 |
DeepSeek says off-peak rates are half of peak rates. Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday, excluding Chinese public holidays. The pricing page also warns that product prices may vary and recommends checking the page regularly for the most recent pricing. DeepSeek Models & Pricing
The practical takeaway is simple: V4 Pro should not be your default if cost and throughput matter most. It has higher token prices and a lower concurrency limit than deepseek-flash. Use it only when it produces better results for your exact workload.
DeepSeek V4 Pro vs DeepSeek V4.1 Flash
The best current comparison is deepseek-v4-pro versus deepseek-flash. V4 Pro is the higher-cost Pro route. V4.1 Flash is the newer Flash route and the better default for most new cost-sensitive API projects. For workload selection and cost scenarios, see our DeepSeek Flash vs V4 Pro comparison.
| Comparison point | deepseek-v4-pro | deepseek-flash |
|---|---|---|
| Current model version | DeepSeek-V4-Pro-0813. | DeepSeek-V4.1-Flash. |
| Best default for new projects | No, use only after testing. | Yes, for most cost-sensitive hosted API work. |
| Vision | Not supported. | Supported. |
| Context length | 1M. | 1M. |
| Maximum output | 384K. | 384K. |
| Thinking Mode | Supported. | Supported. |
| JSON Output / Tool Calls / Responses API | Supported. | Supported. |
| Output price, off-peak | $1.98 / 1M tokens. | $0.60 / 1M tokens. |
| Output price, peak | $3.96 / 1M tokens. | $1.20 / 1M tokens. |
| Concurrency limit | 500. | 2500. |
| Lifecycle nuance | Phaseout/routing was announced, then service continuation appeared in the Change Log. | Current Flash route; old V4 Flash names route to it for compatibility. |
Use deepseek-flash first when you need low cost, Vision, higher concurrency, or a general default model. Test deepseek-v4-pro when your workload involves harder reasoning, code-agent behavior, or technical tasks where higher cost may be justified by better results.
DeepSeek V4 Pro vs DeepSeek R1 and deepseek-reasoner
DeepSeek V4 Pro is not DeepSeek R1. R1 is a historically important open-weight reasoning model family, while V4 Pro is part of the V4 family and is currently represented in the API model table as deepseek-v4-pro. Old R1-era tutorials that use deepseek-reasoner should not be copied into new V4 Pro integrations without checking the current DeepSeek docs.
DeepSeek’s April 24, 2026 V4 changelog entry said the old legacy API names deepseek-chat and deepseek-reasoner would be discontinued after July 24, 2026, and that during the transition they pointed to non-thinking and thinking modes of deepseek-v4-flash. That makes those names poor defaults for new V4 Pro content. DeepSeek API Change Log
Open weights and self-hosting
The official Hugging Face model card for DeepSeek-V4-Pro-0813 lists the license as MIT. It also provides usage paths for Transformers, vLLM, SGLang, Docker Model Runner, inference providers, notebooks, and local apps. DeepSeek-V4-Pro-0813 model card
That does not mean self-hosting is free or easy. The model card includes local and server deployment paths, but operating a large model requires compute, memory, storage, monitoring, security, and inference expertise. Third-party inference providers may also use different pricing, context limits, output limits, or feature support than the official DeepSeek API.
If you want the easiest path, use the hosted API. If you need model-weight access, deployment control, or local testing, review the model card, license, hardware requirements, and serving stack carefully before choosing V4 Pro weights.
When to use DeepSeek V4 Pro
DeepSeek V4 Pro is most relevant when you have a workload where Pro behavior beats Flash enough to justify the cost. Do not choose it only because the name sounds stronger.
| Use case | Why V4 Pro may fit | What to verify |
|---|---|---|
| Hard technical reasoning | Pro may be worth testing when reasoning quality matters more than cost. | Compare against deepseek-flash on your own prompts. |
| High-effort coding agents | The model card emphasizes improved agentic capabilities and production-environment performance. | Measure accepted patches, tests passed, latency, and review time. |
| Tool-heavy workflows | Tool Calls and Responses API are supported in the current model table. | Validate tool-call arguments and total cost. |
| Long-context technical work | The current table lists a 1M context window and 384K max output. | Evaluate accuracy over long context, not just context size. |
| Open-weight Pro experiments | The Hugging Face model card lists weights and MIT license. | Check serving stack, infrastructure cost, and security controls. |
| Existing V4 Pro production workflows | Keeping V4 Pro may reduce migration risk if it performs well. | Reconfirm current API status, pricing, and Change Log before relying on it. |
When not to use DeepSeek V4 Pro
Do not use V4 Pro by default if a cheaper model works. Consider deepseek-flash first when:
- You need the lowest hosted API cost.
- You need Vision or image input.
- You need higher concurrency.
- You are building high-volume extraction, classification, or summarization pipelines.
- Your workload is output-heavy and V4 Pro’s output cost would dominate spend.
- You cannot tolerate lifecycle ambiguity around V4 Pro status without rechecking the docs.
- Your benchmark shows
deepseek-flashperforms similarly for the task. - You are following an old tutorial that uses
deepseek-chat,deepseek-reasoner, or V4 Pro Preview wording.
Common mistakes from old tutorials
Many DeepSeek tutorials were written before the V4 Pro GA release, V4.1 Flash, or the later service-continuation note. These mistakes can lead to wrong model choices, higher costs, or broken examples.
| Old mistake | Why it is a problem | Use this instead |
|---|---|---|
| Treating V4 Pro Preview as the current model. | DeepSeek-V4-Pro-0813 superseded the preview version. | Refer to DeepSeek-V4-Pro-0813 for current V4 Pro information. |
| Saying V4 Pro is definitely retired. | The Change Log and pricing table currently show continued V4 Pro service/listing. | Explain the phaseout announcement and later continuation note. |
| Saying V4 Pro definitely routes to V4.1 Flash. | That was announced in the V4.1 Flash release, but later docs show a more nuanced current state. | Tell readers to check the current Models & Pricing page. |
Using deepseek-reasoner as the default model. | It is old R1-era guidance and not the current V4 Pro model ID. | Use deepseek-v4-pro or deepseek-flash according to the current model table. |
Using deepseek-v4-flash as the current Flash default. | It is a legacy compatibility name routed to V4.1 Flash. | Use deepseek-flash. |
| Assuming V4 Pro supports images. | The current model table says Vision is not supported for deepseek-v4-pro. | Use deepseek-flash for Vision. |
| Ignoring output cost. | V4 Pro output tokens are much more expensive than Flash output tokens. | Estimate cache-hit input, cache-miss input, and output separately. |
Developer checklist before using V4 Pro in production
- Check the current DeepSeek Models & Pricing page.
- Check the DeepSeek API Change Log for lifecycle updates.
- Confirm that
deepseek-v4-prois still listed for your intended use. - Compare against
deepseek-flashon the same prompts and tasks. - Measure quality, latency, retry rate, output length, and total cost per successful task.
- Do not use V4 Pro for image input; use
deepseek-flashif Vision is required. - Track cache-hit input, cache-miss input, and output tokens separately.
- Keep model names in configuration rather than hard-coding them in many files.
- Do not copy old
deepseek-chatordeepseek-reasonerexamples into new projects. - If self-hosting, review the model card, license, hardware, serving stack, monitoring, and security requirements.
Frequently asked questions
What is DeepSeek V4 Pro?
DeepSeek V4 Pro is the Pro route of DeepSeek’s V4 model family. The current hosted API model name is deepseek-v4-pro, and the current model table maps it to DeepSeek-V4-Pro-0813.
Is DeepSeek V4 Pro still available?
The current DeepSeek Models & Pricing page still lists deepseek-v4-pro, and the API Change Log says DeepSeek decided to continue providing API service for V4 Pro after September 14, 2026. Always check the live docs before deployment.
Does deepseek-v4-pro still work?
Current official documentation still lists deepseek-v4-pro. But because DeepSeek previously announced a phaseout/routing plan, production users should verify the current model table and Change Log before relying on it.
Did DeepSeek retire V4 Pro?
DeepSeek announced a V4 Pro phaseout/routing plan in the V4.1 Flash announcement, but the later/current Change Log says DeepSeek decided to continue V4 Pro API service after September 14, 2026. Do not describe it as simply retired without checking the current docs.
Did DeepSeek reverse the V4 Pro shutdown?
DeepSeek’s Change Log says that in response to user demand, DeepSeek decided to continue providing API service for V4 Pro after September 14, 2026, with billing unchanged. That is the key update to mention alongside the earlier V4.1 Flash phaseout announcement.
What is DeepSeek-V4-Pro-0813?
DeepSeek-V4-Pro-0813 is the official GA release of DeepSeek V4 Pro. The Hugging Face model card says it supersedes the preview version and improves agentic capabilities and production-environment performance.
Is DeepSeek V4 Pro better than V4.1 Flash?
Not universally. V4 Pro may be worth testing for harder tasks, but V4.1 Flash is cheaper, supports Vision, and has a higher concurrency limit. Use your own evaluation rather than assuming one is always better.
Should I use deepseek-v4-pro or deepseek-flash?
Start with deepseek-flash for most new cost-sensitive projects. Use deepseek-v4-pro only if your workload shows a measurable quality improvement that justifies the higher price and lower concurrency limit.
Does DeepSeek V4 Pro support image input?
No. The current Models & Pricing page lists Vision as not supported for deepseek-v4-pro. Use deepseek-flash if you need image input through the official DeepSeek API.
Does DeepSeek V4 Pro support Thinking Mode?
Yes. The current model table says deepseek-v4-pro supports both thinking and non-thinking modes, with thinking as default.
Does DeepSeek V4 Pro support Tool Calls?
Yes. Tool Calls are listed as supported for deepseek-v4-pro in the current DeepSeek model table.
Does DeepSeek V4 Pro support JSON Output?
Yes. JSON Output is listed as supported for deepseek-v4-pro.
What is the context length of DeepSeek V4 Pro?
The current DeepSeek model table lists a 1M-token context length for deepseek-v4-pro.
What is the max output of DeepSeek V4 Pro?
The current DeepSeek model table lists a 384K maximum output for deepseek-v4-pro.
How much does DeepSeek V4 Pro cost?
In the current official pricing table, deepseek-v4-pro costs $0.022 per 1M cache-hit input tokens, $0.66 per 1M cache-miss input tokens, and $1.98 per 1M output tokens off-peak. Peak prices are $0.044, $1.32, and $3.96 respectively. Check the current page before production use because pricing can change.
Is DeepSeek V4 Pro open source?
Use precise wording. The official Hugging Face model card lists DeepSeek-V4-Pro-0813 under the MIT license and provides model weights. Hosted API use is still billed separately by DeepSeek.
Can I self-host DeepSeek V4 Pro?
The Hugging Face model card provides local and serving paths such as Transformers, vLLM, SGLang, and Docker Model Runner. But self-hosting a large model requires serious compute, storage, security, monitoring, and engineering work.
Is deepseek-reasoner still the model to use?
No, not as the default for V4 Pro. deepseek-reasoner belongs to old R1-era guidance. For current V4 Pro, check the Models & Pricing page and use current model IDs such as deepseek-v4-pro or deepseek-flash.
Has V4.1-Pro launched?
The checked official sources mention V4.1-Pro as a future route in the V4.1 Flash announcement, but the current Models & Pricing page lists deepseek-flash and deepseek-v4-pro. Do not claim V4.1-Pro has launched unless the current official docs say so.
Bottom line
DeepSeek V4 Pro is best understood as the Pro route of the V4 family, currently represented in the official model table by deepseek-v4-pro and DeepSeek-V4-Pro-0813. It is not the same as R1, deepseek-reasoner, V4 Flash, or deepseek-flash.
The safest recommendation is to start with deepseek-flash for most new DeepSeek API projects because it is cheaper, supports Vision, and has a higher concurrency limit. Use deepseek-v4-pro when your own evaluation shows that Pro behavior is worth the higher price, lower concurrency, and lack of Vision support.
The main editorial point is nuance: DeepSeek did announce a V4 Pro phaseout/routing plan in the V4.1 Flash release, but the current Change Log and model table show continued V4 Pro service/listing. Any article about DeepSeek V4 Pro should explain that lifecycle clearly instead of simply calling the model “retired” or “unchanged.”