DeepSeek-V4.1-Flash (Max): Task Cost and Net Improvement in Agent Arena

One-sentence takeaway

Arena.ai reports a +4.87% net improvement for DeepSeek-V4.1-Flash (Max) relative to the Arena baseline. The body of the post states a $0.07 median cost per task, while the detailed table states $0.06. The cost should therefore be recorded as an inconsistency in the original post, rather than combined into one unconditionally certain figure.

Use cases

• Tasks this can help assess: Tool orchestration, task completion, and controllability in real Agent Mode tasks, as well as the tradeoff between cost and net improvement.

• Tasks this should not be extrapolated to: General chat, single-fact knowledge questions, standalone coding problems, visual tasks, or other tasks outside Agent Arena; it also cannot be used to infer the per-call cost of every invocation.

• Applicable model version: DeepSeek-V4.1-Flash (Max); the original post does not provide an API ID, so this article alone cannot justify relabeling the historical model tag as another version.

• Test environment or client: Agent Arena's Agent Mode; the leaderboard page says the data comes from real Agent Mode tasks.

• Reasoning tier and parameters: DeepSeek uses Max; the comparison includes Max, High, and xHigh. Temperature, context, tool configuration, and other parameters are not specified.

Evaluation method

The original post uses Agent Arena's Pareto frontier and net improvement relative to the Arena baseline to report each model's net improvement percentage and median cost per task. It does not disclose the task sample size, task prompts, tool set, number of runs, aggregation formula, or margin of error.

The linked Agent Arena page (collected on 2026-09-16) defines the chart as “Net Improvement × Cost/task from real Agent Mode tasks.” It uses 50th percentile for the cost statistic and labels it “Median cost per task (last 14 days).” On that date, the page showed 1,850,083 sessions and 46 models, indicating that this was a dynamic leaderboard snapshot. It showed Deepseek V4.1 Flash (Max) at +4.88%, $0.06/task, a normal snapshot change from the previous day's +4.87% in the X post.

The original post does not define the Arena baseline's specific score, sample, or calculation method. The page only summarizes its signals as including tool reliability, task completion, and controllability. “Net improvement” can therefore only be understood as an internal Arena leaderboard metric relative to its baseline, not as general accuracy or win rate.

Key results

• The original post says DeepSeek-V4.1-Flash (Max) has the lowest median cost per task among the top three open-source models.

• Relative to Hy4 preview: retains 98% of its net improvement at 73% lower cost.

• Relative to Kimi K3 (Max): retains 76% of its performance at 92% lower cost.

• Relative to Fable 5 or stronger models: retains 35–54% of their net improvement at 97–99% lower cost; these models cost 37–76 times more per task.

• The original post also says that after this release, GPT-5.6 Luna (xHigh), GLM-5.3-Flash, and DeepSeek-V4-Flash left the Agent Arena Pareto frontier. This is a change at the leaderboard boundary and does not mean these models became less capable on all tasks.

Raw data

The detailed title in the original post is “Net improvement over Arena baseline | Median cost/task.” The figures below preserve the original post's numbers; see “Conclusions and limitations” for the inconsistency in the cost figures.

The cost relationships are expressions rounded in the original post. For example, using the table's $0.06, $0.06/$0.22≈27%, meaning the cost is about 73% lower than Hy4; $0.06/$0.77≈8%, meaning the cost is about 92% lower than Kimi K3.

Conclusions and limitations

The supported conclusion is that, for real Agent Mode tasks in Arena and using the median cost over roughly the past 14 days, DeepSeek-V4.1-Flash (Max) achieves +4.87% net improvement at approximately $0.06–$0.07/task and sits on the cost-efficiency Pareto frontier. It is suitable for cost-performance comparisons when selecting an Agent model.

The limitations are:

• The body's $0.07 and the detailed table's $0.06 do not match; the original post does not explain whether this is due to rounding, different statistical timestamps, or a display error.

• The original post does not disclose the Arena baseline, task sample, task distribution, tool configuration, cost breakdown, confidence interval, or significance test.

• The Agent Arena leaderboard is a dynamic snapshot; the page showed +4.88% and $0.06/task the following day, so leaderboard figures should not be treated as permanently fixed results.

• “Net improvement” is not general accuracy; “cost reduction” applies only to the median per-task cost metric used by the same leaderboard.

Reproduction notes

Open the original post in Tabbit and click “Show original” to read the English content, avoiding errors in the direction of the multiplier caused by automatic translation. Then open the Agent Arena Pareto page linked from the same post and check 50th percentile, last 14 days, and the dynamic snapshot date. Because the original post does not provide the task set, prompts, tools, or parameters, the +4.87% net improvement cannot be independently reproduced; only the publicly displayed leaderboard snapshot can be cross-checked.