By dssub · Reviewed November 9, 2026
Why is the V4.1 Pro question back?
Developers have been watching for the next Pro model since DeepSeek introduced V4.1 Flash. A recent report from Max For AI, a LobeHub developer and maintainer, adds a useful data point: on a private set of agent tasks, the current V4 Pro experience on DeepSeek’s website reportedly completed more tasks consistently than the earlier V4 Pro-0813 result.
The evaluator suggested a possible limited test of V4.1 Pro. That is a hypothesis, not confirmation of the model’s identity. In the official materials reviewed on November 9, we found no announced V4.1 Pro release date.
What did the 50-task test measure?
LobeBench Hard 50 is a private collection of complex agent tasks. The evaluator says it tests whether an agent can understand a request, use tools and keep working until the task is complete. Each task was attempted three times.
| Reported result | V4 Pro-0813 | Current V4 Pro on DeepSeek’s website |
|---|---|---|
| Passed all three attempts | 24 of 50 · 48% | 32 of 50 · 64% |
| Passed at least once | 80% | 82% |
The all-three result rose by eight tasks, or 16 percentage points. The at-least-once rate moved by two percentage points. In this test, the reported difference was therefore more pronounced on repeated completion than on whether a task succeeded at least once.
These are the evaluator’s reported results, not a dssub test. The task set is private, and the full methods and runs have not been independently reproduced here.
Does the result prove a new model is live?
No. A score measures the behavior of the tested setup; it does not identify the underlying model checkpoint. The model, routing, tool environment, prompts and other configuration may all affect an agent’s results. Without more information, the benchmark cannot isolate which factor caused the difference.
The result also concerns DeepSeek’s website. It should not be assumed to describe every request made through the official API or an independent provider. DeepSeek’s API changelog says the company continued providing V4 Pro API service after September 14, superseding the routing plan described in its earlier Flash announcement.
What should developers take from the numbers?
A single success and three consecutive successes answer different questions. If an agent occasionally solves a difficult task, it may be useful when a person can review and retry the result. For a workflow that needs dependable unattended execution, consistency matters more. Neither measure alone guarantees production performance.
Fifty tasks and three attempts are a limited sample. A private benchmark also makes it harder for others to inspect the task selection or reproduce the comparison. Treat the reported gain as a reason to evaluate your own workflows, not as a universal ranking or proof of a release.
How can you check your own workflow?
- Choose representative tasks. Include both routine jobs and cases that have failed before. Define what a successful result looks like before running them.
- Keep the setup comparable. Record the provider, date, model ID, client and version, prompts, tools and settings. Preserve any model version or fingerprint the provider returns.
- Repeat each task. Record every attempt, including failures, rather than keeping only the best result. Track completion, errors, elapsed time and usage.
- Review the pattern. Compare the share of tasks completed at least once with the share completed on every attempt. Look at failure types before changing a production workflow.
These are suggested evaluation steps, not a test performed by dssub. A model name that remains unchanged does not by itself prove the underlying service has remained unchanged.
What is available through dssub?
dssub is an independent service and is not affiliated with DeepSeek. Its current connection documentation lists deepseek-flash and deepseek-v4-pro; the benchmark above does not establish the version or performance of a request made through dssub. Check your account’s connection details and test the tasks that matter to you.
Explore your setup options
Review models and API compatibility or compare the DeepSeek Subscription and their usage limits.