Eight More Tasks Completed in a 50 Task Test
DeepSeek's current V4 Pro model completed 32 of 50 tasks in an internal benchmark run by MaxForAI, an AI educator and developer and maintainer of the LobeHub agent client. The earlier V4 Pro-0813 version completed 24. The results, described in accounts dated November 8, 2026, put the newer model at 64%, up from 48%.
Contents
The most interesting improvement was consistency. The two versions performed similarly when measured by whether they could finish a task at least once. The larger gap appeared when the test required successful completion three times in succession, a demanding measure for software expected to carry out work reliably.
The findings have fueled speculation that DeepSeek is gradually testing V4.1 Pro through its existing service. The model's identity and release schedule remain unconfirmed, and the benchmark results were presented under the current V4 Pro name.
The testing discussion followed a partial service interruption on November 7, the second reported disruption within a week. DeepSeek said it had implemented a fix and was monitoring the results. For customers, the immediate developments therefore concern two different needs: better task completion and dependable access to the service.
What Changed in LobeBench?
MaxForAI used LobeBench, an internal test collection, to compare the current V4 Pro with V4 Pro-0813. The older version placed 13th in the reported ranking. The newer version ranked first among Chinese models in that comparison.
The increase was 16 percentage points, with eight additional tasks classified as Hard completed successfully. Measured against the earlier total of 24 completed tasks, the increase to 32 amounts to about one third more successful tasks. Both calculations describe the same result from different angles: a higher share of the test completed and a larger number of tasks finished.
LobeBench's repeated completion requirement helps explain why the result attracted attention. A model that succeeds once but fails on subsequent attempts can be difficult to use for routine work. An agent client, which lets an AI system carry out tasks through software, needs dependable execution as well as the ability to produce an occasional strong answer.
The reported comparison centers on 50 tasks and an internal ranking. Details needed to reproduce it, including the complete task list, evaluation rules, model settings and scores for competing systems, were not published in the descriptions of the results.
Why Repeated Success Matters
An AI model can generate different responses to the same request. In a task involving several steps, an early mistake can also affect everything that follows. Requiring three successful attempts in succession puts more weight on dependable behavior than counting any single successful attempt.
MaxForAI's comparison reportedly found little difference between the two versions when either could receive credit for succeeding at least once. The gap widened under the repeated success measure. That makes reliability the central finding, rather than simply a new ability to solve tasks that the earlier version could never finish.
For someone using an agent to complete work, this distinction has practical value. Repeated failures can mean more retries, more checking and more human intervention. A model that completes the same task consistently may be more useful even when an occasional successful answer from an older version looks equally good.
Other informal demonstrations circulating around the suspected update included animation rendering and an image of a cormorant riding a bicycle. LobeBench offered a numerical comparison against a named earlier version, shifting attention toward repeatable task performance.
What Happened During the Outages?
DeepSeek's November 7 interruption began at 15:11, according to the company's status information described at the time. The affected services included chat, the V4 Pro API and the V4.1 Flash API. An API is the connection that allows another application to send requests to DeepSeek automatically.
Users encountered an unavailable chat function and a message saying the server was busy. DeepSeek later said a fix had been implemented and that it was monitoring the result. Subsequent testing described the service as working normally again.
A similar incident had occurred on November 3, beginning at 16:58 and affecting the same three services. After a fix was announced for that incident, the services were still listed as having reduced performance. Restoring access and restoring normal speed can happen at different stages of an incident.
The two interruptions form part of the backdrop to the model speculation. The cause of the November 7 outage was not specified in the status information described, leaving customers without a confirmed explanation connecting it to either traffic growth or model deployment.
A Timeline of the Pro Expectations
Interest in V4.1 Pro predates the latest test results. September release plans, October changes to the service and the November benchmark comparison have each contributed to expectations, although they concern different parts of DeepSeek's product line.
- September 10, 2026: V4.1 Flash was released. A proposed plan to route requests for deepseek-v4-pro to Flash until V4.1 Pro arrived was subsequently withdrawn following developer objections.
- October 17: The V4.1 Flash-1017 version was released, according to the October account of subsequent service changes.
- October 27: Observers detected changes in model fingerprints, starting around 18:21. By 19:07, the fingerprints observed across three model options had converged to one.
- November 3: A partial interruption affected chat and both named API services, starting at 16:58.
- November 7: Another partial interruption began at 15:11. DeepSeek later announced a fix.
- November 8: Accounts of MaxForAI's LobeBench comparison described the increase from 24 to 32 completed tasks.
The September routing proposal matters to developers because an application's model name can be part of its operating assumptions. Changing the model behind that name can affect responses, costs and performance. After the proposal was withdrawn, V4 Pro continued to be offered with unchanged billing.
What the October Fingerprints Revealed
The October 27 observations concerned model fingerprints, identifiers that can help track changes in the system serving a request. The old fingerprint was associated with V4.1 Flash-1017. The replacement identifier could not be interpreted well enough to identify the model behind it.
At the time, the observed change appeared to concern Flash rather than Pro. Possible explanations included a different model checkpoint, meaning a saved version of a model, or a fresh deployment of the same model. The new identifier left that choice unresolved.
V4.1 Flash-1017 had been available for only 10 days when those changes appeared. Expectations about Pro were already strong, making even a change associated with Flash a focus of attention. The November LobeBench result is more directly relevant to Pro because it compares two versions carrying the V4 Pro label.
DeepSeek's earlier release timing also shaped expectations. The V4 generation reportedly had a gap of about two weeks between Flash and Pro. By the later discussion of V4.1, the wait after Flash had exceeded seven weeks. The current release cycle has therefore followed a different timetable.
What Developers Can Compare Today
Reported pricing places V4 Pro and V4.1 Flash at different cost levels. V4 Pro input was listed at $0.66 per million tokens, with output outside peak periods at $1.98 per million tokens. Flash input and output were listed at $0.15 and $0.60 per million tokens respectively.
Tokens are the small pieces of text a model processes, such as words or parts of words. At those listed rates, Pro input costs 4.4 times as much as Flash input. The quoted Pro output rate outside peak periods is 3.3 times the listed Flash output rate. Actual spending depends on the amount of input and output, timing and applicable caching.
Flash also had a reported cache hit price of $0.003 per million tokens. A cache lets the service reuse previously processed input. Its listed capabilities included native image input, allowing images to be passed directly to the model.
V4 Pro's reported capacity included a context window of one million tokens and an output limit of 384,000 tokens. The context window is the amount of material the model can consider in a request. These specifications describe the existing product against which customers would compare a future Pro release.
What Would Make the Update Official?
A release announcement naming V4.1 Pro would give users a definite product to evaluate. A matching API documentation entry, model repository and pricing update would provide practical details such as how to access it, what it costs and which capabilities it supports.
A review of DeepSeek's public release information described the latest relevant repository as DeepSeek-V4.1-Flash, created on September 10, 2026. It also said the 51 page technical report accompanying Flash contained neither the name V4.1-Pro nor V4.1 Pro. The public product information described in that review still centered on Flash and V4 Pro.
Optimistic speculation around the November benchmark results suggested a release during the following week. DeepSeek had provided no public date, release window or quarter in the release information described. For now, the concrete development is the reported improvement in repeated task completion under the current V4 Pro name.
Key Points
- V4 Pro completed 32 of 50 LobeBench tasks, compared with 24 for V4 Pro-0813.
- The score rose from 48% to 64%, an increase of 16 percentage points.
- The main improvement appeared when tasks required three successful completions in succession.
- DeepSeek experienced partial service interruptions on November 3 and November 7, affecting chat and two API services.
- The October fingerprint changes appeared to concern Flash, while the November benchmark comparison concerned Pro.
- V4.1 Pro's identity, release date and pricing remain unconfirmed.
