DeepSeek V4.1 Flash Takes First Place in 9 Composite Rankings

Benchmark update

One model, the same 60 exercises

DeepSeek V4.1 Flash completed 33 Mathematical Analysis III and 27 General Physics II exercises from the 2025 and 2026 exams. Each exercise was submitted separately with the same prompt and dataset used for the existing panel. Thinking was enabled with high reasoning effort, matching the earlier V4 Flash reasoning configuration.

Answers were evaluated blind against the approved, frozen T1–T8 rubrics before unblinding. The comparison now covers 17 model-mode configurations. V4.1 Flash does not set a new accuracy record; its main result is the balance among correctness, response time and estimated cost.

Composite rankings

The nine rankings it leads

V4.1 Flash ranks first under Quality first, Balanced and Low cost in each of the three subject selections. These are nine preset-and-subject combinations—not nine independent quality criteria or nine separate experiments.

Quality first assigns 75% to correctness, 15% to speed and 10% to cost. Balanced uses 50%, 25% and 25%; Low cost uses 30%, 15% and 55%. All three assign zero weight to token consumption. Their scores combine normalized metrics through the site’s weighted geometric formula; they are not raw rubric scores.

Data & methodOpen the current Balanced ranking ↗.

Correctness

High correctness, without the accuracy lead

Removing resource trade-offs changes the order. Under Accuracy only, V4.1 Flash is fourth overall, third in Analysis III and sixth in Physics II. Sol leads overall; the earlier V4 Flash is immediately ahead of V4.1 Flash.

V4.1 Flash and V4 Flash have almost identical overall means: the unrounded difference is below 0.01 rubric points. V4.1 scores higher in Analysis III but lower in Physics II. It produces 32 perfect answers, compared with 42 for V4 Flash, while retaining a median score of 100.

Compared with V4 Pro, V4.1 gains 0.52 mean rubric points, although Pro retains the higher minimum score. Sol leads V4.1 by 1.49 points overall. These are observed benchmark differences, not evidence of universal superiority.

Resource profile

Response time explains much of the change

V4.1 Flash’s median elapsed time is 51.0% lower than V4 Flash’s and 54.4% lower than V4 Pro’s. It is also 70.0% lower than Sol’s, at an estimated cost 94.4% lower. That combination lets it lead Balanced even though Sol remains more accurate.

Within DeepSeek, the cost picture is different: V4.1 Flash costs about 4.06 times as much as the archived V4 Flash runs and 71.4% more than V4 Pro. It wins on the weighted balance, not by being the cheapest. Its billed output is only 4.5% lower than V4 Flash’s, and its total-token consumption is the second-highest in the panel.

Data & methodResource values come from the run-derived token-and-cost CSV ↗.

Limits

Where the lead stops

V4.1 Flash is fourth overall under Low latency, where speed receives 60% of the weight. Under Equal weights, which gives token consumption the same importance as correctness, speed and cost, it drops to fourteenth overall. High token use matters once that metric enters the composite.

Correctness is strong but not uniform: two answers score below 80, including a 50-point Physics II response. Independent verification reaches 99.29 where applicable, while final results are the lowest aggregate rubric dimension at 96.22. The result is strong without implying that final-answer errors have disappeared.