<think>**Translating product availability phrase** </think> DeepSeek V4 Pro is no longer being discontinued.

<think>Planning faithful markdown translation

</think>

DeepSeek Withdraws Its Decommissioning Plan; V4 Pro API Will Remain Available

DeepSeek confirmed today that the V4 Pro API service, originally scheduled to go offline at 12:00 PM Beijing Time on September 14, 2026, will continue to accept requests after that time, with the existing pricing structure unchanged.

This means that developers already using deepseek-v4-pro in production will temporarily not need to change the model name or migrate prompts, nor worry about requests being switched to another model without explicit notice. DeepSeek said in its announcement that it decided to continue offering the V4 Pro API in response to user demand. If there are any subsequent changes, it will issue a separate notice.

The key point of this adjustment is not that the model’s capabilities have suddenly changed, but that DeepSeek has temporarily returned control over the model lifecycle to its users.

A Three-Day Reversal Over the Model’s Decommissioning

DeepSeek’s previous handling of V4 Pro was not complicated, but the pace was extremely fast—nearly a different version of the plan each day.

• September 9: DeepSeek announced that before V4.1 Pro went online, all user requests to V4 Pro would be routed to V4.1 Flash and billed at V4.1 Flash’s prices.
• September 10: DeepSeek further clarified that after 12:00 PM Beijing Time on September 14, it would officially discontinue the V4 Pro service, with V4.1 Flash taking over the requests.
• September 10: V4.1 Flash officially launched. DeepSeek said that after extensive testing across performance, cost, speed, total execution time, and other metrics, the model outperformed V4 Pro overall.
• September 11: DeepSeek announced that it would continue providing the V4 Pro API after September 14, with the existing pricing structure unchanged.

From the model provider’s perspective, the original plan was reasonable: if a newer, cheaper, and more capable Flash model is already available, continuing to maintain an older Pro service increases the costs of routing, compute scheduling, and version maintenance. For many ordinary workloads that do not depend on fixed model behavior, automatic migration could also directly reduce users’ bills.

For developers, however, a model name has never been merely a label.

An online agent, code review workflow, or batch-generation task is often tuned around a specific model’s output style, tool-calling patterns, latency range, and context behavior. Even if a new model performs better on public benchmarks, that does not mean it can seamlessly replace the old model in every business workflow. Routing deepseek-v4-pro directly to V4.1 Flash is, in essence, the service provider performing an unconfirmed version upgrade on the user’s behalf.

This was the core of the earlier controversy: developers were not necessarily opposed to V4.1 Flash; rather, they objected to their production systems’ model dependencies being changed unilaterally.

Pricing Remains Unchanged, While the Price Gap Between V4 Pro and Flash Is Still Significant

According to the prices currently published, DeepSeek uses a pricing model that distinguishes between cache hits and cache misses, as well as peak and off-peak periods. Off-peak prices are half of peak prices. Peak hours are 9:00 AM–12:00 PM and 2:00 PM–6:00 PM, Monday through Friday, Beijing Time; all other times are considered off-peak.

| Model | Input price (cache hit, off-peak/peak) | Input price (cache miss, off-peak/peak) | Output price (off-peak/peak) | | --- | ---: | ---: | ---: | | DeepSeek V4.1 Flash | RMB 0.02 / RMB 0.04 | RMB 1 / RMB 2 | RMB 4 / RMB 8 | | DeepSeek V4 Pro | RMB 0.15 / RMB 0.30 | RMB 4.5 / RMB 9 | RMB 13.5 / RMB 27 |

All prices above are calculated per million tokens.

On paper, V4.1 Flash has a clear pricing advantage. Taking output charges for cache misses as an example, V4 Pro costs RMB 27 per million tokens during peak hours, while V4.1 Flash costs RMB 8. If a business primarily handles large-scale summarization, batch classification, ordinary code generation, or back-office automation, Flash is clearly better suited for cost control.

But “cheaper” does not automatically mean “should replace the existing model.” For some high-value tasks, stability, established output formats, tool-calling success rates, and regression results may matter more than saving a few yuan per million tokens. In particular, for teams that have already built evaluation sets and monitoring systems, the cost of switching models often lies not in the API bill, but in retesting and fixing edge cases in production.

Therefore, the most direct significance of this announcement for developers is: V4 Pro users can continue using the original model as before, while Flash users can continue to enjoy lower prices—neither group is forced into an either-or choice.

V4.1 Flash Is More Capable, but Version Differences Cannot Be Erased

Released by DeepSeek on September 10, V4.1 Flash is the smaller model in its new model architecture series and natively supports multimodal visual understanding. According to the company, it outperformed flagship models, including V4 Pro, on multiple benchmarks.

This shows that competition in the model market no longer fully follows the product-naming logic that “Pro must be stronger than Flash.” With advances in training, inference, and model architecture optimization, smaller models can also outperform larger, older models in speed, price, and certain agent tasks.

For practical development, V4.1 Flash is better suited to the following scenarios:

1. High-concurrency, low-cost tasks: such as initial customer-service screening, content classification, batch rewriting, and log analysis.
2. Interactive applications that require fast responses: Flash is specifically positioned toward throughput and response speed.
3. Large-scale experimentation with agents and tool calls: when the commercial value of a large number of tasks has not yet been proven, a low-cost model is better able to absorb the cost of exploration.
4. Workflows requiring visual input: V4.1 Flash natively supports multimodal visual understanding and can handle image-related tasks.

V4 Pro, meanwhile, still has a clearly defined user base: production systems that have already been tuned around it through prompt optimization, output-structure constraints, and business regression testing. For these users, retaining the original model is more important than chasing benchmark rankings.

It is important to note that DeepSeek’s decision to “continue providing” V4 Pro does not mean that V4 Pro has entered a long-term stable maintenance phase. The announcement only states that the service will continue after September 14. It does not provide a new end date, nor does it promise that prices, rate limits, or routing will never change in the future. Teams should therefore not interpret this extension as a guarantee that the model will never change.

What Developers Should Do Now

If your business is currently calling V4 Pro, it is advisable to do three things first.

1. Record the Actual Model Version Reliably

Do not record only the model name included in the request. Logs should also preserve the model information returned in the response, request time, prompt version, tool-calling results, and error rate. This way, even if routing changes occur on the server side in the future, you can determine whether business performance was affected by a model upgrade or by some other infrastructure change.

2. Establish a Comparative Evaluation of V4 Pro and V4.1 Flash

Do not look only at public leaderboards. At a minimum, use your own real-world samples to compare the following metrics:

• Time to first token and total response time;
• Input and output tokens and actual cost;
• Compliance rate for JSON or structured output;
• Tool-call success rate and parameter error rate;
• Information retrieval performance in long-context tasks;
• Test pass rate after code generation;
• Stability of instruction following in multi-turn conversations.

If evaluation results show that Flash is already good enough, you can migrate proactively. If V4 Pro is more stable in critical workflows, there are also sufficient reasons to continue using it. The key is to let business metrics—not model names—make the decision.

3. Reserve a Configuration Switch for Model Migration

Even if you are still using V4 Pro today, it is advisable to place the model name, pricing strategy, and routing logic in a configuration center rather than hard-coding them into business code. When V4.1 Pro launches or V4 Pro enters another decommissioning process, your team can validate the change through a gradual rollout instead of having to modify the entire request chain at short notice.

This Reversal Exposes a Fundamental Issue with API Services

For chat applications, model upgrades are often product-experience improvements. For API users, however, the model version is part of the infrastructure.

Developers build an entire set of implicit contracts around a model: context window, output format, tool calls, refusal boundaries, reasoning patterns, latency, and price. A provider may believe that “the new model is stronger,” but users need predictability. Especially in payment, code deployment, data processing, and enterprise automation scenarios, unconfirmed routing changes may create greater risks than differences in model capabilities.

This is the most noteworthy aspect of the announcement: DeepSeek’s ultimate decision to retain V4 Pro is not merely a service extension; it is also a response to the practical need to prioritize compatibility in API products.

Of course, continuing to provide V4 Pro will also bring additional compute and operational pressure. Now that V4.1 Flash is online, DeepSeek will likely continue encouraging users to migrate, but this time the migration will need clearer version descriptions, a transition period, and greater freedom of choice for developers. For model providers, the speed of model iteration is important; for API platforms, a stable contract is equally important.

For now, V4 Pro remains available, with pricing unchanged. V4.1 Flash offers lower costs, higher throughput, and native multimodal capabilities. The more prudent strategy is not to bet on one model or the other, but to place both within the same evaluation and routing system: let Flash handle scalable, cost-sensitive tasks while V4 Pro continues serving production workflows that have already been validated.

If you access DeepSeek or other mainstream models through OpenAI Hub, you can also use its unified OpenAI-compatible interface for model comparisons and gradual rollout testing. Before migrating, however, you should still rely on actual response results, cost statistics, and business regression data rather than looking only at the model name or official benchmarks.

Conclusion

DeepSeek has confirmed that the V4 Pro API will not be decommissioned as originally planned after September 14. It will remain available for requests, with pricing unchanged. The previous plan to automatically route V4 Pro requests to V4.1 Flash is, at least for now, no longer an outcome developers are required to accept.

In the short term, this is a more user-friendly decision for existing customers. In the long term, what really matters is whether DeepSeek can make future model upgrades predictable, gradual, and reversible through proper API version management. V4.1 Flash’s performance and pricing advantages are certainly worth testing, but whether it should replace V4 Pro should be determined by developers’ own production data.